Presentation on theme: "Component based SEM Comparison between various methods"— Presentation transcript:
1 Component based SEM Comparison between various methods Michel Tenenhaus
2 A Component-based SEM tree ALL BLOCK REFLECTIVESEMComponent-based SEM(Score computation)Covariance-based SEM (CSA)(Model estimation)Chatelin-Esposito VinziFahmy-Jäger-TenenhausXLSTAT-PLSPM (2007)W. ChinPLS-GraphHerman WoldNIPALS (1966)PLS approach (1975)J.-B. LohmöllerLVPLS 1.8 (1984)H. HwangY. TakaneGSCA (2004)VisualGSCA 1.0(2007)(AMOS 6.0, 2007)Score computedusing block MVloadingsPath analysis on the structural modeldefined on the scoresFor good blocks(High Cronbach ):- Score = 1st PC- Score = MV’sPath-PCAULS-SEMGSCAPath-ScalePLSWhen the blocks are heterogeneous,GSCA is too close to PCA. PLS and SEMgive almost the same results.M. Tenenhaus : Component-based SEMTotal Quality Management, 2008When all blocks are good, all the methodsgive almost the same results.
6 the LVs coming from the 5 methods Comparison betweenthe LVs coming from the 5 methodsPCAULS-SEMSCALEPLSGSCAWhen all blocks are good, all the methodsgive almost the same results.
7 ECSI model with noise Noise variables are highly correlated (> .99) and uncorrelated with Customer Satisfaction MVs.For this new block:- Noise = 1st PC- Customer Satisfaction = 2nd PC
8 Fornell weights when the augmented Customer Satisfaction block is heterogeneous and reflective GSCA istrapped !!!!
9 Why GSCA is trapped The GSCA criterion PCA MSEV, Glang (1988) MSEV = Maximum Sum of Explained Variance
10 For reflective blocks, GSCA seems to be too close to PCA Fornell weightsfor originalECSI model
11 Fornell weights when the augmented Customer Satisfaction block is heterogeneous and formative GSCA is stilltrapped !!!!
12 = A Component-based SEM tree ALL BLOCK FORMATIVE Component-based SEM Herman WoldPLS approach (1975)Mathes (1994)Component-based SEM(Score computation)H. HwangVisualGSCA 1.0(2007)M. GlangMSEV (1988)=Glang and Hwang criteria are equivalent.Computational practice: PLS MaximumPLS Critical points
17 Comparison between methods ****Criterion optimized by the methodPractice supports “theory”
18 the LVs coming from the 3 methods Comparison betweenthe LVs coming from the 3 methodsB + CentroidB + FactorialGSCAWhen all blocks are good, all the methodsgive almost the same results.
19 Economic inequality and political instability (Russet) Agricultural inequality (X1)GINIINST+++ECKS+FARM1+++DEATRENT-3D-STB+GNPR++D-INS-2LABO-DICTIndustrialdevelopment (X2)Politicalinstability (X3)
20 Use of XLSTAT-PLSPM Mode B + Centroid scheme Y1 = X1w1Y3 = X3w3Y2 = X2w2
21 Use of XLSTAT-PLSPM Mode B + Factorial scheme Y1 = X1w1Y3 = X3w3Y2 = X2w2
22 Use of GSCA (All formative) When there is only one structural equation andwhen all blocks are formative,GSCA is equivalentto a canonical correlation analysis.
23 Use of XLSTAT-PLSPM for two blocks Mode B Canonical Correlation Analysis
24 Comparison between methods *****Criterion optimized by the methodPractice supports “theory”
25 ConclusionWhen the blocks are good (or moderately good) all methods seems to give almost the same LV scores.When some blocks are heterogeneous, PLS and ULS-SEM seems to give better results than GSCA.For all formative blocks : GSCA criterion is a more natural criterion than the PLS ones.For all formative blocks : PLS give good results for multiblock data analysis.
26 Final conclusion William Camden (1623) « All the proofs of a pudding are in the eating, not in the cooking ».William Camden (1623)