A bridge between PLS path modeling and ULS-SEM Michel Tenenhaus 1 A SEM tree SEM Component-based SEM (Score computation) Herman Wold NIPALS (1966) PLS approach (1975) Svante Wold Harald Martens PLS regression (1983) SIMCA-P The Unscrambler Covariance-based SEM (CSA) (Model validation) H. Hwang Y. Takane Generalized Structured GSCA (2004) K. Joreskog (LISREL, 1970) Component Analysis (ALS) J.-B. Lohmöller LVPLS 1.8 (1984) W. Chin PLS-Graph H. Hwang VisualGSCA 1.0 (2007) C. Ringle SMART-PLS Chatelin-Esposito Vinzi Fahmy-Jäger-Tenenhaus XLSTAT-PLSPM (2008) R. McDonald(1996) M. Tenenhaus (2001) 2 Covariance-based Structural Equation Modeling Latent Variables : 1 η m 1 ξ k Endogenous LV Exogenous LV Structural model (Inner model): η = Bη + Γξ + ζ 3 Structural Equation Modeling Measurement model (outer model) : y j1 jy1 j1 yj = j y y jp jp j j jp j λ yj y = Λy η + ε MV LV Endogenous x 1 x1 1 x = x p xp p λx εj δ x = Λ xξ + δ MV LV Exogenous 4 Structural Equation Modeling Mixing Structural and measurement models : 1 - η = Bη + Γξ + ζ η = (I - B) (Γξ + ζ) - x = Λ xξ + δ 1 - y = Λy η + ε y = Λy (I - B) (Γξ + ζ) + ε = Cov() = E(’) = Cov() = E(’) = Cov() = E(’) = Cov() = E(’) Residual variances are diagonal matrices 5 Structural Equation Modeling Covariance matrix for manifest variables : Outer model Inner model Exogen. LV Cov. Structural Residual variance Σ xx Σ( Λ x , Λ y , B, Γ, Φ, Ψ, Θε , Θδ ) Σ yx Measurement Residual variance Σ xy Σ yy 1 Λ xΦΛ x ' + Θδ Λ xΦΓ ' (I - B) ' Λ y ' 1 1 1 Λ y (I - B) ΓΦΛ x ' Λ y [(I - B) (ΓΦΓ' + Ψ )][(I - B ) '] Λ y ' + Θ ε 6 Covariance-based SEM ULS algorithm : S = Observed covariance matrix for MV Λ x , Λ y , B, Γ,Φ, Ψ, Θε , Θδ Minimize PCA Generalization S Σ( Λ x , Λ y , B, Γ, Φ, Ψ, Θε , Θδ ) 2 Λx , Λy ,B,Γ ,Φ, Ψ ,Θε ,Θδ Goodness-of-fit Index (Jöreskog & Sorbum): GFI 1 ˆ ,Λ ˆ , Bˆ , Γ,Φ, ˆ ,Θ ˆ ,Θ ˆ ) ˆ ˆ Ψ S Σ( Λ x y ε δ S 2 2 7 Path model describing causes and consequences of Customer Satisfaction: The inner model Image Loyalty Customer Expectation Perceived value Customer satisfaction . Perceived quality Complaints 8 Use of AMOS 6.0 e2 e1 1 1 e5 e4 e3 1 1 e6 1 PQ2 CE1 CE2 CE3 1 e8 1 PQ3 PQ1 e7 PQ4 e9 1 1 PQ5 PQ6 e10 1 1 1 PQ7 CUST_EXP Méthod = ULS 1 PER_QUALI d1 1 e11 e12 1 1 PV1 PV2 d5 1 PER_VALUE 1 1 ima1 e19 e20 d2 1 1 d7 ima2 d3 1 1 1 ima3 e21 CUS_SAT IMAGE 1 e22 1 e13 CSI1 ima4 1 First Roderick McDonald’s idea (1996) 1 e23 ima5 Measurement residual variances are canceled: ˆ =Θ ˆ =0 Θ ε δ 1 e14 CSI2 1 CSI3 e15 1 d4 CUST_ LOY COMPLAINTS 1 0.8 This is a computational trick: Residual variances are passed to errors and can always be computed afterwards. CL1 1 e16 CL2 1 e17 CL3 1 e18 1 Complaints e24 9 Covariance-based SEM ULS algorithm with the McDonald’s constraints: S = Observed covariance matrix for MV Ω ( Λ x , Λ y , B, Γ, Φ, Ψ, Θε 0, Θδ 0) 2 Minimize S Σ( Λ x , Λ y , B, Γ, Φ, Ψ , Θε 0, Θδ 0) Λ x , Λ y ,B, Γ ,Φ , Ψ Outer model Inner model Goodness-of-fit Index (Jöreskog & Sorbum): GFI 1 ˆ ,Λ ˆ ,B ˆ ,Θ ˆ ˆ ˆ , Γ,Φ, ˆ ˆ Ψ S Σ( Λ x y ε , final , Θδ , final ) S 2 2 10 0 Use of AMOS 6.0 1 0 e1 CE2 e4 1 CE3 CE1 1 PQ2 PQ3 PQ4 1 0 1 CUST_EXP e8 0 0 PER_QUALI 0 1 1 PQ5 1 PQ1 - Méthod = ULS - Measurement residual variances = 0 1 1 0 1 e7 e6 e5 e3 0 0 0 0 e2 1 e12 e11 1 PQ6 1 1 e9 d1 d5 PV2 PV1 PQ7 1 0 1 e10 0 1 PER_VALUE ima1 e19 1 0 1 1 d2 ima2 e20 d3 1 0 1 1 e21 0 1 CSI1 e13 CUS_SAT ima3 IMAGE 0 e22 0 1 1 e14 CSI2 ima4 0 1 e23 0 ima5 CSI3 1 e15 1 d4 CUST_ LOY Complaints 1 1 d6 CL1 1 0 e16 CL2 1 0 e17 CL3 1 0 e18 11 .00 Results 1 .00 e1 e4 1 GFI = .869 CE1 CE3 1.03 1.11 1.00 PQ2 1 Outer LV Estimates: 1 .00 1 .98 ima3 1.18 .00 1 e10 18.51 d3 1 196.90 .40 .00 1 1.00 CSI1 ima4 e13 CUS_SAT IMAGE 1.53 .00 1 1 .89 1.68 .20 .00 1 e23 PQ6 .57 .79 .00 e9 d2 ima2 1 e22 PQ7 1 285.53 .15 1.00 .00 e21 54.73 1.33 1 PER_VALUE 1.00 CSI1 1.53 CSI2 1.68 CSI3 CSI 1 1.53 1.68 e20 .00 1.04 1.00 .95 .59 e8 d1 .51 -.60 1 PQ5 PV2 PV1 ima1 e19 1 e12 .68 2nd McDonald’s idea + Fornell’s idea .00 .00 .93 1 1 d5 PQ4 .00PER_QUALI e11 29.10 PQ3 1.00 1.06 .00 1 1 .99 1.29 1.03 PQ1 CUST_EXP 1 1 .00 1 CE2 e7 e6 e5 e3 .00 .00 .00 .00 e2 1.40 e14 CSI2 1.52 .00 ima5 CSI3 325.29 1 e15 -.02 Complaints PLS estimate of LV: 1 1.00 .95 .31 402.67 - Mode A - LV inner estimate = theoretical LV 1 1 1 .00 .00 - LV inner estimate computation is useless. .00 d4 1 CUST_ LOY d6 CL1 CL2 e16 e17 CL3 e18 12 Cross-validation by bootstrap PQ2 PQ1 CE1 PQ3 PQ4 1 1.03 CE2 1.11 Expectation CE3 1.29 .99 1 1.03 Perceived quality 1.06 PV1 .95 1 .59 .51 .57 Perceived value 1 PQ7 d1 -.59 IM1 d2 R2=.36 IM3 1.53 .98 Satisfaction .40 .79 R2=.82 Image .59 IM5 CS2 1.68 1 CS3 d3 1.18 IM4 CS1 1 .15 1 IM2 PQ6 1.33 1 PV2 d5 .68 1.04 R2=.67 R2=.70 1 PQ5 .93 .61 .89 Loyalty 1 d4 -.02 Complaints R2=.44 1 .31 R2=.37 1 .95 d6 CL1 CL2 CL3 13 Figure 2 : Estimation of the ECSI model (Significant links in bold, non-significant dotted) Cross-validation by bootstrap Parameter Customer expectation <--- Image Perceived quality <--- Customer expectation Perceived value <--- Perceived quality Perceived value <--- Customer expectation Customer satisfaction <--- Perceived value Customer satisfaction <--- Customer expectation Customer satisfaction <--- Image Customer satisfaction <--- Perceived quality x6 (Complaints) <--- Customer satisfaction Customer loyalty <--- Customer satisfaction Customer loyalty <--- Image Customer loyalty <--- x6 (Complaints) x11 <--- Image x12 <--- Image x13 <--- Image x14 <--- Image x15 <--- Image x21 <--- Customer expectation x22 <--- Customer expectation x23 <--- Customer expectation x31 <--- Perceived quality x32 <--- Perceived quality x33 <--- Perceived quality x34 <--- Perceived quality x35 <--- Perceived quality x36 <--- Perceived quality x37 <--- Perceived quality x41 <--- Perceived value x42 <--- Perceived value x51 <--- Customer satisfaction x52 <--- Customer satisfaction x53 <--- Customer satisfaction x71 <--- Customer loyalty x72 <--- Customer loyalty x73 <--- Customer loyalty Estimate 0.593 1.058 0.511 0.679 0.152 -0.595 0.402 0.568 1.519 1.405 0.201 -0.018 1 0.792 0.981 1.184 0.890 1 1.034 1.112 1 0.989 1.295 1.031 0.929 1.043 1.328 1 0.954 1 1.535 1.682 1 0.312 0.953 Inf (95%) Sup (95%) P 0.380 0.807 0.050 0.777 1.638 0.050 -0.195 0.979 0.170 0.000 1.883 0.100 0.059 0.246 0.050 -1.538 0.048 0.070 0.184 0.918 0.050 0.158 0.847 0.060 1.089 1.938 0.050 0.806 5.161 0.050 -2.296 0.558 0.520 -0.165 0.120 0.790 1 1 ... 0.553 1.057 0.05 0.652 1.527 0.05 0.972 1.551 0.05 0.718 1.151 0.05 1 1 ... 0.559 1.589 0.05 0.675 1.693 0.05 1 1 ... 0.736 1.255 0.05 1.069 1.547 0.05 0.881 1.207 0.05 0.773 1.102 0.05 0.900 1.223 0.05 1.082 1.606 0.05 1 1 ... 0.851 1.083 0.05 1 1 ... 1.269 1.899 0.05 1.294 2.139 0.05 1 1 ... -0.069 0.662 0.11 0.808 1.127 14 0.05 Comparison between the Fornell-ULS and Fornell-PLS standardized weights Manifest variable Image Image Image Image Image Customer expectation Customer expectation Customer expectation Perceived quality Perceived quality Perceived quality Perceived quality Perceived quality Perceived quality Perceived quality Perceived value Perceived value Customer satisfaction Customer satisfaction Customer satisfaction Customer loyalty Customer loyalty Customer loyalty Fornell-ULS standardized weight 0.206 0.163 0.202 0.244 0.184 0.318 0.329 0.353 0.131 0.130 0.170 0.135 0.122 0.137 0.174 0.512 0.488 0.237 0.364 0.399 0.441 0.138 0.421 Fornell-PLS standardized weight 0.199 0.173 0.187 0.242 0.198 0.326 0.316 0.357 0.139 0.121 0.168 0.134 0.119 0.135 0.183 0.492 0.508 0.242 0.354 0.404 0.393 0.130 0.478 15 Comparison between the Fornell-ULS and Fornell-PLS standardized weights and all Cor(LVFornell-ULS , LVFornell-PLS) > .99 16 First particular case : Factor Analysis and Principal Component Analysis 17 a posteriori computation First particular case : FA and PCA .00 .02 1 Capacity 1 Capacity e1 .00 .14 1 Power .99 1.00 F1 e2 1 Speed .87 e3 1.00 Weight 1 F1 e4 e2 1 Speed .89 .21 e3 .00 .76 Weight 1 .42 e4 .80 .74 .00 .45 .73 Width 1 .80 e5 Width 1 e5 Length Factor Analysis Reflective mode 1 e6 .36 .00 .47 AVE = .69 .15 .00 .92 .53 .68 1 Power .96 .25 .93 .08 e1 Length 1 Principal component Analysis Formative mode AVE = .74 18 e6 .36 FA vs PCA : Variance reconstruction Covariances Cylindrée Cylindrée Puissance Vitesse Poids Largeur Longueur 1 .954 .885 .692 .706 .664 1 .934 1 .529 .466 1 .730 .619 .477 .527 .578 .795 1 .591 1 Puissance Vitesse Poids Largeur Longueur Implied covariances (FA) Cylindrée Puissance Vitesse Poids Cylindrée Puissance Vitesse Poids Largeur Longueur 1 .918 .860 .678 .737 .722 1 .804 1 .633 .593 1 .689 .645 .508 .674 .632 .498 1 .541 1 Largeur Longueur FA does not care for variance reconstruction sij ˆij .169 2 i j Implied covariances (PCA) FA (reflective mode) yields to Largeur better covariances PCA cares for Cylindrée Puissance Vitesse Poids Longueur variance reconstruction (formative mode). Cylindrée .926 .889 than .853PCA .728 .771 .765 Puissance Vitesse Poids Largeur Longueur .853 .818 .785 .699 .671 .573 .740 .710 .606 .734 .705 .602 .642 .637 .632 reconstruction s i j ij ˆ ij 19 .229 2 MIMIC mode (with this new approach) (Multiple effect indicators for multiple causes) .09 .00 1 e4 Capacity Weight 1.00 .00 1 e5 .81 Width 1 .21 .89 F1 .81 e6 e1 .95 .77 .00 1 1 e2 Power .83 .31 Length Formative mode (multiple cause) Speed 1 e3 Reflective mode (multiple effect) 20 MIMIC mode (usual in CSA) ? .43 d 1.00 -.05 Weight 1 .27 .48 1.00 .59 ? 1.00 Length Formative mode (multiple cause) .09 1 e1 1.00 .07 .55 Width .79 Capacity .94 F1 1 e2 Power .88 .20 Speed 1 e3 Reflective mode (multiple effect) 21 MIMIC mode (usual in CSA) .42 d 1.00 .00 Weight .27 .48 1.00 Width .79 .59 Capacity 1 1.00 Length Formative mode (multiple cause) ? .09 1 e1 1.00 .07 .54 .95 F1 1 e2 Power .89 .19 Speed 1 e3 Reflective mode (multiple effect) 22 MIMIC mode (better) PCA oriented vs the dependent block e6 .46 d .00 .00 1 Weight 1 .84 .00 e5 e4 e1 1.00 .07 1.00 1 .87 Width 1 .95 .75 F1 e2 Power F2 .88 .00 1 1 Capacity .89 .19 Length Speed Proposal: Compute a global score as i Cov( X i , F2 ) * X i / Formative mode (multiple cause) i 1 e3 Cov( X i , F2 ) Reflective mode (multiple effect) (Same than before) 23 Second particular case : Multi-block data analysis 24 3 Appellations 4 Soils Sensory analysis of 21 Loire Red Wines (J. Pagès) 4 blocks of variables 2el (Saumur),1 X1 X2 X3 X4 Smell intensity at rest 3.07 Aromatic quality at rest 3.00 Fruity note at rest 2.71 Floral note at rest 2.28 Spicy note at rest 1.96 Visual intensity 4.32 Shading (orange to purple) 4.00 Surface impression 3.27 Smell intensity after shaking 3.41 Smell quality after shaking 3.31 Fruity note after shaking 2.88 Floral note after shaking 2.32 Spicy note after shaking 1.84 Vegetable note after shaking 2.00 Phenolic note after shaking 1.65 Aromatic intensity in mouth 3.26 Aromatic persisitence in mouth 3.26 Aromatic quality in mouth 3.26 Intensity of attack 2.96 Acidity 2.11 Astringency 2.43 Alcohol 2.50 Balance (Acid., Astr., Alco.) 3.25 Mellowness 2.73 Bitterness 1.93 Ending intensity in mouth 2.86 Harmony Illustrative3.14 Global quality 3.39 variable 1cha (Saumur),1 1fon (Bourgueil),1 1vau (Chinon),3 … 2.96 2.82 2.38 2.28 1.68 3.22 3.00 2.81 3.37 3.00 2.56 2.44 1.74 2.00 1.38 2.96 2.96 2.96 3.04 2.11 2.18 2.65 2.93 2.50 1.93 2.89 2.96 3.21 2.86 2.93 2.56 1.96 2.08 3.54 3.39 3.00 3.25 2.93 2.77 2.19 2.25 1.75 1.25 3.08 3.08 3.08 3.22 2.18 2.25 2.64 3.32 2.68 2.00 3.07 3.14 3.54 2.81 2.59 2.42 1.91 2.16 2.89 2.79 2.54 3.16 2.88 2.39 2.08 2.17 2.30 1.48 2.54 2.54 2.54 2.70 3.18 2.18 2.50 2.33 1.68 1.96 2.46 2.04 2.46 … … … … … … … … … … … … … … … … … … … … … … … … … … … … t1 (Saumur),4 t2 (Saumur),4 3.70 3.19 2.83 1.83 2.38 4.32 4.00 3.33 3.74 3.08 2.83 1.77 2.44 2.29 1.57 3.44 3.44 3.44 2.96 2.41 2.64 2.96 2.57 2.07 2.22 3.04 2.74 2.64 3.71 2.93 2.52 2.04 2.67 4.32 4.11 3.26 3.73 2.88 2.60 2.08 2.61 2.17 1.65 3.10 3.10 3.10 3.33 2.57 2.67 2.70 2.77 2.31 2.67 3.33 3.00 2.85 X1 = Smell at rest, X2 = View, X3 = Smell after shaking, X4 = Tasting PCA of each block: Correlation loadings SMELL AT REST VIEW SMELL AFTER SHAKING TASTING 1.0 1.0 Spicy note 0.8 Vegetable note 0.6 T2 0.8 Smell intensity Phelonic note T1 3EL Aromatic persistency in mouth 0.4 Aromatic intensity mouth 0.2 PER1 4EL 1VAU 1TUR -0.0 -0.2 2ING 1FON 1CHA 1POY 2DAM 1BOI Fruity note Smell quality Aromatic quality in mouth Floral note 2ING -0.0 0.2 0.4 Harmony Balance -0.6 -1.0 -0.2 Intensity of attack -0.4 -1.0 -0.4 4EL 2BOU 2BEA 1ING PER1 1ROC 1DAM 1BOI 1TUR 3EL DOM1 1BEN 2DAM 1POY 1CHA 1FON 2EL -0.2 -0.8 -0.6 Ending intensity in mouth 0.2 -0.8 -0.8 T1 1VAU Mellowness -0.4 -1.0 Astringency Alcohol T2 0.4 -0.0 1DAM Bitterness 0.6 2BEA in DOM1 1ING 2EL 1ROC 2BOU 1BEN -0.6 Acidity 0.6 0.8 1.0 -1.0 -0.8 -0.6 -0.4 -0.2 -0.0 0.2 0.4 0.6 0.8 1.0 PCA of each block: Correlation loadings e1 1 .00 e20 1 .85.91 rest4 .34 .00 view3 View shaking1 1 shaking2 .47 .88 1 shaking3 .82 .33 1 e9 .00 1 e8 .00 1 e21 .00 shaking4 shaking6 1.00 Shaking 1 .09 -.64 .37 .89 shaking5 .88 .89 .00 shaking10 shaking7 1 .96 shaking8 .00 e24 1 shaking9 e23 1.00 tasting7 .90 .84 .77 tasting1 1 .00 tasting2 1 .00 tasting8 .77 1 e18 .00 e17 e26 tasting4 1 .00 e16 .00 1 e25 .00 1 tasting6 tasting5 tasting3 .00 .38 .94 e19 1 .97 Tasting 1 -.26 e27 tasting9 1 .00 1 e22 .98 .99 1.00 .95 .74 1 e12 .00 e10 .00 1 view1 view2 rest1 Rest 1 e13 .00 .00 1 1 1.00 .08 rest5 e7 e6 1 1 rset2 .00 .00 e5 e4 1 rest3 .00 .00 e3 e2 1 .00 e11 GFI = .301 .00 .00 e14 .00 1 e15 Multi-block data analysis = Confirmatory Factor Analysis .00 .00 e1 1 .00 SMELL AT REST e20 1 .77.90 rest4 .50 .00 .00 .98 .87 shaking1 .83 1 shaking2 .59 .79 .72 .23 1 shaking4 shaking6 .74 1.00 Shaking 1 .25 -.54 .37 .88 shaking5 .89 .76 .91 e24 1 shaking9 e23 1.00 tasting7 .85 .80 .77 tasting1 1 .00 tasting2 1 .00 tasting8 .77 1 e18 .00 e17 .00 1 e25 .00 1 tasting6 tasting5 tasting3 .00 e26 .35 .85 e19 1 .92 Tasting 1 -.20 e27 .91 shaking8 .00 1 tasting9 1 .00 1 .00 .92 shaking10 shaking7 e22 VIEW .89 .91 1.00 View Rest 1 shaking3 e9 .00 1 e8 .00 1 e21 .00 view1 view2 .72 1 e10 .00 1 GFI = .849 view3 rest1 1 e12 .00 e11 1 1 1.00 .00 rest5 1 .59 e7 e6 1 rset2 .00 .00 e5 e4 1 rest3 e13 .00 SMELL AFTER SHAKING e3 e2 1 .00 .00 .00 e14 TASTING .00 1 e15 tasting4 1 .00 e16 28 First dimension .00 .00 .00 e5 e3 e2 .84 .94 view1 view2 .90 .99 .88 .65 1.00 1.00 View Rest 1 Using MV with significant loadings 1 1 view3 rest2 rest3 e7 e6 1 1 1 .00 .00 .87 .00 .79 1 e12 shaking2 .84 .00 1 shaking3 e11 .69 1.00 .78 Shaking 1 .00 .87 1 shaking8 e22 .88 .84 .81 .00 1 e23 .90 shaking9 .00 e24 1 1.00 shaking10 .93 Tasting 1 tasting9 e27 .91 .85 .89 1 tasting8 .00 e26 .84 .76 .00 1 tasting1 tasting6 1 1 .00 e19 tasting4 1 .00 e16 tasting5 1 .00 e15 .00 e14 29 First global score .88 e7 e6 view1 view2 1.00 1.12.98 1.00 .00 .00 1 d1 Rest 1 1 View .79 d2 .80 1 e12 shaking2 1.00 1.04 .00 1 .96 shaking3 e11 GFI = .973 1 1 view3 rest2 rest3 .00 .00 1 1 1 .00 e5 e3 e2 2nd order CFA .00 .00 .00 Score 1 .78 Shaking 1 .00 1.08 1 1 shaking8 e22 1.09 .00 d3 1.00 .00 .89 1 e23 .88*rest3 + 1*rest2 Rest 1 .88 + 1 shaking9 .00 e24 1 shaking10 .00 1 d4 1.00 Tasting 1 e27 .99 .91 .95 1 tasting8 Score 1 .79*Rest 1 + .80*View + .78*Shaking 1 + .89*Tasting 1 .79 + .80 + .78 + .89 .00 e26 .91 .82 .00 1 tasting9 tasting1 tasting6 1 1 .00 e19 tasting4 1 .00 e16 tasting5 1 .00 e15 .00 e14 30 Validation of the first dimension Correlations Rest1 View Shaking1 Tasting1 Score1 Rest1 1 .621 .865 .682 .813 View 1 .762 .813 .920 Shaking1 Tasting1 1 .895 .942 1 .944 31 .00 .00 e20 1 e4 Second dimension 1 rest5 rest1 .74 .91 1.00 Rest 2 .00 1 e13 .75 shaking1 .77 .90 1.00 .00 .74 1 Shaking 2 shaking5 e9 .57 .00 e21 .79 1 shaking7 1.00 Tasting 2 .87 tasting3 tasting7 1 1 .00 e17 .81 .00 e25 32 .00 .00 e20 1 e4 2nd global score 1 rest5 rest1 1.00 1.25 .00 1 Rest 2 d1 GFI = .905 .66 .00 1 e13 .00 shaking1 1 .00 e9 1.34 1 1.00 d3 1.62 Score 2 .52 Shaking 2 shaking5 .75 1.00 .00 Score 2 1 e21 shaking7 .66*Rest 2 + .52*Shaking 2 + .75*Tasting 2 .66 + .52 + .75 .00 Tasting 2 1.08 tasting3 d4 1.00 tasting7 1 1 .00 e17 1 .00 e25 33 Validation of the second dimension Correlations Rest2 Rest2 Shaking2 Tasting2 Score2 1 .789 .782 .944 Shaking2 Tasting2 1 .803 .904 1 .928 34 Mapping of the correlations with the global scores Score 2 unrelated with quality Score 1 related with quality 35 Correlation with global quality Variables related to dimension 1 Aromatic quality at rest Fruity note at rest Visual intensity Shading (from orange to purple) Surface impression Smell quality Aromatic intensity in mouth Aromatic persistence in mouth Aromatic quality in mouth Intensity of attack Alcohol Balance (acidity, astringency, alcohol) Mellowness Ending intensity in mouth Harmony Variables related to dimension 2 Smell intensity at rest Spicy note at rest Smell intensity after shaking Spicy note after shaking Phelonic note Astringency Bitterness Global quality 0.62 0.50 0.54 0.51 0.67 0.76 0.61 0.68 0.85 0.77 0.52 0.95 0.92 0.80 0.88 New result. Not obtained with other multi-block data analysis methods, nor with factor analysis of the whole data. Global quality 0.04 -0.31 0.17 -0.08 0.09 0.41 0.05 36 Wine visualization in the global score space Wines marked by Appellation 37 Wine visualization in the global score space Wines marked by Soil 38 Visualization of wine variability among the blocks Star-plot of the “best wine” – 2DAM SAUMUR 3,0 DAM = Dampierre-sur-Loire 2,8 2,6 2,4 2DAM GLOBAL SCORE Tasting 2,2 Smell after shaking View 2,0 2,25 Smell at rest 2,50 2,75 3,00 3,25 3,50 Cuvée Lisagathe 1995 A soft, warm, blackberry nose. A good core of fruit on the palate with quite well worked tannin and acidity on the finish; Good length and a lot of potential. DECANTER (mai 1997) (DECANTER AWARD ***** : Outstanding quality, a virtually perfect example) Third particular case : Analysis of covariance between two blocks of binary variables (with C. Guinot et E. Mauger (CERIES)) Data = Sun exposure = Sun protection A = Gender (A1 = Men, A2 = Women) Model 1 < 0? 2 < 3 ? = 0 + 1A1 + 2 *A1 + 3 *A2 + (1) 42 Theory: background = 0 + 1A1 + 2 *A1 + 3 *A2 + W (1) W M No gender effect 1 = 0, 2 = 3 M Gender main effect 1 0, 2 = 3 Interaction *gender 2 3 43 Theory: background X1 Sun exposure during lifetime (4) Score for sun exposure X2 Sun exposure during mountain sports (2) X3 Sun exposure during nautical sports (2) ˆ Xw X4 Sun exposure during hobbies (2) X5 Practice of naturism (1) Y Sun protection behavior over the past year (6) Score for sun protection ˆ Yc A Gender 44 Theory: background = 0 + 1A1 + 2 *A1 + 3 *A2 + (1) Equation (1) is replaced by: Yc = 0 + 1A1 + 2Xw*A1 + 3Xw*A2 + (2) = 0 + 1A1 + 2(X*A1)w + 3(X*A2)w + (3) Question: How to estimate and test the parameters w, c, 0 , 1, 2, 3 ? 45 Theory: methods Covariance based SEM with constraints X1*A1 X2*A1 w1 w2 w3 X*A1 2 Y X3*A1 3 X1*A2 w1 X2*A2 w2 w3 X3*A2 c1 c2 c3 1 Y1 Y2 Y3 A1 X*A2 No group effect on the measurement model 46 Application: material We have applied the ULS-SEM to a study on sunexposure behavior in 8,084 French adults. Development of skin cancers Premature skin ageing. Data came from the SU.VI.MAX study* *Hercberg S. et al. Arch Intern Med. 2004;164:2325-42 47 0 1 e11 e10 0 e9 0 1 s15_1_h 1 e7 w1 s28_34_h 1 1 1 1 s25j2_h 0 1 e5 s26_h 0 e4 1 s26j2_h 0 s28_34_f w4 s29_34_f Sun exposure (W) 1 w7 w7 W8 w8 1 1 W10 e1 1 0 e28 1 s27j2_f e21 0 1 s21_f e22 M en Sun Protection 1 1 0 e20 1 d s6_1 e19 0 1 s27_f s21_h e18 0 1 s26j2_f s27j2_h e150 0 1 1 0 0 e17 s26_f w11 e14 e160 1 s25j2_f w9 w9 w10 w11 1 s25_f s27_h 0 e2 w2 w3 w5 Sun Exposure (M ) e13 0 1 e3 1 s15_1_f w5 s25_h e12 0 w4 0 e6 s14z3_f w1 w2 w3 s29_34_h e8 0 Use of AMOS method = ULS s14z3_h 0 0 1 c2 s6quand2 1 0 e27 c3 s6regul c5 c4 s6vis c6 s6cps 1 0 1 0 1 0 e26 e25 e24 s6hors1 1 0 e23 48 . Sun exposure of body and face Sun exposure between 11am and 4 pm Basking in the sun is important or very important Intensity of sun exposure moderate or severe Sun exposure during practice of mountain sports Nb of days of mountain sport actvities > 200 days AMOS results 2.82 2.82 GFI = .870 2.30 . . 1.46 3.91 3.91 2.13 1.00 Sun expo. (M) 2.13 Sun expo. (W) 1.00 1.23 .64 Men 2.67 -.17 .62 .62 [-.20, -.14] d Several times while sun exposure Sun exposure during practice of nautical sports Nb of days of practice of nautical activites > 400 days Sun exposure during practice of hobbies Nb of days of lifetime hobbies > 900 days .62 Product used for face with SPF > 15 .40 .82 While sun exposure Nb of days of mountain sport actvities > 200 days Practice of naturism during lifetime Sun Protection While sun tanning Sun exposure between 11am and 4 pm Basking in the sun is important or very important Intensity of sun exposure moderate or severe Sun exposure during practice of mountain sports 1.35 1.38] 1.24 [1.12, .75 1.00 . 2.30 1.46 Sun exposure during practice 1.23 of nautical sports Confidence Interval Nb of days of practice of nautical .64 activites > 400(Bootstrap) days 2.67 Sun exposure during practice of hobbies[.64, .89] 1.35 Nb of days of lifetime hobbies > 900 days Practice of naturism during lifetime Sun exposure of body and face .90 .45 Product used for body with SPF > 15 Product used besides voluntarily sun exposure periods Results: AMOS ULS = 0 + 1M + 2*M + 3*W + Coefficients Estimate Lower Upper 2 Sun exposure (Men) Sun Protection .754 .647 891 3 Sun exposure (Women) Sun Protection 1.237 1.123 1.385 1 Men Sun Protection -.166 -.198 -.139 Conclusion 1. Women tend to protect themselves from the sun more than men (1 < 0). 2. This difference between men and women increases as lifetime sun exposure increases (3 - 2 > 0) 50 LV estimation using PLS (Mode A, Fornell’s normalisation) ˆ LV ( X ) w1 X 1 w2 X 2 w11 X 11 w X j j w1 w2 w11 c1Y1 c2Y2 c6Y6 ˆ LV (Y ) ck Yk c1 c2 c6 Example for Sun Protection: 1.00 c1 .239 1.00 .82 .90 .62 .40 .45 51 Results Sun protection over the past year score + 0.24 If sun protection products used while sun tanning + 0.20 If sun protection products used throughout voluntarily sun exposure periods + 0.22 If sun protection products applied several times during sun exposure periods + 0.14 If the sun protection product used for the face has a SPF* over 15 + 0.09 If the sun protection product used for the body has a SPF* over 15 + 0.11 If sun protection products used besides voluntarily sun exposure periods *SPF: Sun protection factor 52 Results Lifetime sun exposure score 0.14 If sun exposure of the body and the face 0.11 If sun exposure between 11 a.m. and 4 p.m. 0.07 If basking in the sun is declared important or extremely important 0.20 If self-assessed intensity of sun exposure is declared moderate or severe 0.10 If sun exposure during practice of mountain sports 0.05 If the number of days of lifetime mountain sports activities > 200 days* 0.06 If sun exposure during practice of nautical sports 0.03 If the number of days of lifetime nautical sports activities > 400 days* 0.13 If sun exposure during practice of hobbies 0.07 If the number of days of lifetime hobby activities > 900 days* 0.03 If practice of naturism during lifetime * Median value of the duration was used as a threshold for dichotomisation 53 Results: analysis of covariance ˆ Protection on ˆ Sun exposure and Gender Parameter Intercept score_x1_protect GENRE GENRE score_x1_protec*GENRE score_x1_protec*GENRE Estimate Femmes Hommes Femmes Hommes 0.0729460737 0.2473795070 0.1269948620 0.0000000000 0.1613712617 0.0000000000 B B B B B B Standard Error t Value Pr > |t| 0.01213456 0.02574722 0.01557730 . 0.03316612 . 6.01 9.61 8.15 . 4.87 . <.0001 <.0001 <.0001 . <.0001 . Coefficients Estimate Lower Upper 2 Sun exposure (Men) Sun Protection .754 .647 .891 3 Sun exposure (Women) Sun Protection 1.237 1.123 1.385 1 Men Sun Protection -.166 -.198 -.139 Main effect Gender + Interaction Sun exposure*Gender 54 are highly significant. Conclusion 1: SEM-ULS > PLS • When mode A is chosen, outer LV estimates using Covariance-based SEM (ULS or ML) or Component based SEM (PLS) are always very close. • It is possible to mimic PLS with a covariance-based SEM software (McDonald,1996, Tenenhaus, 2001). • Covariance-based SEM authorizes to implement constraints on the model parameters. This is impossible with PLS. 55 Conclusion 2: PLS > SEM-ULS • When SEM-ULS does not converge or does not give an admissible solution, PLS is an attractive alternative. • PLS offers many optimization criterions for the LV search (but rigorous proofs are still to be found). • PLS still works when the number of MV is very high and the number of cases very small. • The new software XLSTAT-PLSPM will be available at the beginning of 2008. 56 References - Tenenhaus M., Esposito Vinzi V., Chatelin Y.-M., Lauro C. (2005) : « PLS path modeling » Computational Statistics & Data Analysis, 48, 159-205. - M. Tenenhaus, E. Mauger, C. Guinot : « Use of ULS-SEM and PLS-SEM to measure interaction effect in a regression model relating two blocks of binary variables » in Handbook of Partial Least Squares (PLS): Concepts, Methods and Applications (V. Esposito Vinzi, J. Henseler, W. Chin, H. Wang, Eds), Springer, 2007. - M. Tenenhaus : « A bridge between PLS path modeling and ULS-SEM » PLS’07, Oslo. 57 Final conclusion « All the proofs of a pudding are in the eating, not in the cooking ». William Camden (1623)