KF Qualitätsmanagement Vertiefungskurs V Messung und statistische Analyse von Kundenzufriedenheit Outline Customer satisfaction measurement The Structural Equation Model (SEM) Estimation of SEMs Evaluation of SEMs Practice of SEM-Analysis 3.12.2004 Messung & Analyse von Kundenzufriedenheit 2 The ACSI Model Ref.: http://www.theacsi.org/model.htm 3.12.2004 Messung & Analyse von Kundenzufriedenheit 3 ACSI-Model: Latent Variables Customer Expectations: combine customers’ experiences and information about it via media, advertising, salespersons, and word-of-mouth from other customers Perceived Quality: overall quality, reliability, the extent to which a product/service meets the customer’s needs Customer Satisfaction: overall satisfaction, fulfillment of expectations, comparison with ideal Perceived Value: overall price given quality and overall quality given price Customer Complaints: percentage of respondents who reported a problem Customer Loyalty: likelihood to purchase at various price points 3.12.2004 Messung & Analyse von Kundenzufriedenheit 4 Base line* Q2 1995 Q2 1996 Q2 1997 Q2 1998 Q2 1999 Q2 2000 Q2 2001 Q2 2002 Q2 2003 Q2 2004 % Chan ges % Chang es 79.2 79.8 78.8 78.4 77.9 77.3 79.4 78.7 79.0 79.2 78.3 -1.1% -1.1% Personal Computers 78 75 73 70 71 72 74 71 71 72 74 2.8% -5.1% Apple Computer, Inc. 77 75 76 70 69 72 75 73 73 77 81 5.2% 5.2% Dell Inc. NM NM NM 72 74 76 80 78 76 78 79 1.3% 9.7% Gateway, Inc. NM NM NM NM 76 76 78 73 72 69 74 7.2% -2.6% All Others NM 70 73 72 69 69 68 67 70 69 71 2.9% 1.4% Hewlett-Packard Company – HP 78 80 77 75 72 74 74 73 71 70 71 1.4% -9.0% Hewlett-Packard Company – Compaq 78 77 74 67 72 71 71 69 68 68 69 1.5% -11.5% MANUFACTURING/DURA BLES 3.12.2004 Messung & Analyse von Kundenzufriedenheit 5 The European Customer Satisfaction Index (ECSI) Ref.: http://www.swics.ch/ecsi/index.html 3.12.2004 Messung & Analyse von Kundenzufriedenheit 6 ACSIe-Model for Food Retail 0,33 Perceived Quality Emotional Factor 0,35 0,37 0,36 0,73 0,53 Expectations 3.12.2004 Hackl et al. (2000) (-0,01) Customer Satisfaction 0,34 (0,06) Latent variables and path coefficients 0,34 Loyalty Value Messung & Analyse von Kundenzufriedenheit 7 Austrian Food Retail Market Pilot for an Austrian National CS Index (Zuba, 1997) Data collection: December 1996 by Dr Fessel & GfK (professional market research agency) 839 interviews, 327 complete observations Austria-wide active food retail chains (1996: ~50% from the 10.5 B’EUR market) Billa: well-assorted medium-sized outlets Hofer: limited range at good prices Merkur: large-sized supermarkets with comprehensive range Meinl: top in quality and service 3.12.2004 Messung & Analyse von Kundenzufriedenheit 8 The Data Indicators Latent total expected quality (EGESQ), expected compliance with demands (EANFO), expected shortcomings (EMANG) Expectations (E) total perceived quality (OGESQ), perceived compliance with needs (OANFO), perceived shortcomings (OMANG) Perceived Quality (Q) value for price (VAPRI), price for value (PRIVA) Value (P) total satisfaction (CSTOT), fulfilled expectations (ERWAR), comparison with ideal (IDEAL) Customer Satisfaction (CS) number of oral complaints (NOBES), number of written complaints (NOBRI) Voice (V) repurchase probability (WIEDE), tolerance against pricechange (PRVER) Loyalty (L) 3.12.2004 Messung & Analyse von Kundenzufriedenheit 9 The Emotional Factor Principal component analysis of satisfaction drivers staff (availability, politeness) outlet (make-up, presentation of merchandise, cleanliness) range (freshness and quality, richness) price-value ratio (value for price, price for value) customer orientation (access to outlet, shopping hours, queuing time for checkout, paying modes, price information, sales, availability of sales) identifies (Zuba, 1997) staff, outlet, range: “Emotional factor” price-value ratio: “Value” customer orientation: “Cognitive factor” 3.12.2004 Messung & Analyse von Kundenzufriedenheit 10 Structural Equation Models Combine three concepts Latent variables Pearson (1904), psychometrics Factor analysis model Path analysis Wright (1934), biometrics Technique to analyze systems of relations Simultaneous regression models 3.12.2004 Econometrics Messung & Analyse von Kundenzufriedenheit 11 Customer Satisfaction Is the result of the customer‘s comparison of his/her expectations with his/her experiences has consequences on 3.12.2004 loyalty future profits of the supplier Messung & Analyse von Kundenzufriedenheit 12 Expectation vs. Experience Expectation reflects customers‘ needs offer on the market image of the supplier etc. Experiences include 3.12.2004 perceived performance/quality subjective assessment etc. Messung & Analyse von Kundenzufriedenheit 13 CS-Model: Path Diagram Expectations Customer Satisfaction Perceived Quality 3.12.2004 Loyalty Messung & Analyse von Kundenzufriedenheit 14 A General CS-Model Expectations Voice Customer Satisfaction Perceived Quality Loyalty Profits 3.12.2004 Messung & Analyse von Kundenzufriedenheit 15 CS-Model: Structure EX: expectation PQ: perceived quality CS: customer satisfaction LY: loyalty Recursive structure: triangular form of relations 3.12.2004 to from EX EX PQ CS LY X X 0 X 0 PQ 0 CS 0 0 LY 0 0 Messung & Analyse von Kundenzufriedenheit X 0 16 CS-Model: Equations PQ = a1 + g11EX + z1 CS = a2 + b21PQ + g21EX + z2 LY = a3 + b32CS + z3 Simultaneous equations model in latent variables Exogenous: EX Endogenous: PQ, CS, LY Error terms (noises): z1, z2, z3 3.12.2004 Messung & Analyse von Kundenzufriedenheit 17 Simple Linear Regression Model: Y = a + gX + z Observations: (xi, yi), i=1,…,n Fitted Model: Ŷ = a + cX OLS-estimates a, c: s c , a y - cx xy s sx2 minimize the sum of squared residuals 2 ˆ min i ( yi - yi ) S (a, g) a,g sxy: sample-covariance of X and Y 3.12.2004 Messung & Analyse von Kundenzufriedenheit 18 Criteria of Model Fit R2: coefficient of determination the squared correlation between Y and Ŷ: R2 = ryŷ2 t-Test: Test of H0: g=0 against H1:g≠0 t=c/s.e.(c) s.e.(c): standard error of c F-Test: Test of H0: R2=0 against H1: R2≠0 R2 2 F 1 - R2 n - 2 follows for large n the F-distribution with n-2 and 2 df 3.12.2004 Messung & Analyse von Kundenzufriedenheit 19 Multiple Linear Regression Model: Y = a + X1g1 + ... + Xkgk + z = a + x’g + z Observations: (xi1,…, xik, yi), i=1,…,n In Matrix-Notation: y = a + Xg + z y, z: n-vectors, g: k-vector, X: nxk-matrix Fitted Model: ŷ = a + Xc OLS-estimates a, c: c ( X ' X ) -1 X ' y, a y - c1 x1 - ... - ck xk R2 = ryŷ2 F-Test t-Test 3.12.2004 Messung & Analyse von Kundenzufriedenheit 20 Simultaneous Equations Models A 2-equations model: PQ = a1 + g11EX + z1 CS = a2 + b21PQ + g21EX + z2 In matrix-notation: Y = BY + GX + z with z1 PQ Y , X ( EX ) , z CS z 2 0 0 a1 g 11 path coefficients B , G b 21 0 a 2 g 21 3.12.2004 Messung & Analyse von Kundenzufriedenheit 21 Simultaneous Equations Models Model: Y = BY + GX + z Y, z: m-vectors, B: (mxm)-matrix G: (mxK)-matrix, X: K-vector Some assumptions: z: E(z)=0, Cov(z) = S Exogeneity: Cov(X,z) = 0 Problems: Simultaneous equation bias: OLS-estimates of coefficients are not consistent Identifiability: Can coefficients be consistently estimated? 3.12.2004 Messung & Analyse von Kundenzufriedenheit 22 Path Analytic Model d1 Var(d1) = sEX2 PQ = g11EX + z1 CS = b21PQ + g21EX + z2 EX g21 g11 PQ CS b21 z1 3.12.2004 z2 z 1 s 12 0 Var 2 z 2 0 s 2 Messung & Analyse von Kundenzufriedenheit 23 Path Analysis Wright (1921, 1934) A multivariate technique Model: Variables may be structurally related structurally unrelated, but correlated Decomposition of covariances allows to write covariances as functions of structural parameters Definition of direct and indirect effects 3.12.2004 Messung & Analyse von Kundenzufriedenheit 24 Example d1 sCS,EX = g21s2EX + b21sPQ,EX = g21s2EX + g11b21s2EX EX g11 PQ s YX i(Y X) Yis iX CS b21 z1 3.12.2004 z2 g21 with standardized variable EX: rCS,EX = g21 + g11b21 Messung & Analyse von Kundenzufriedenheit 25 Direct and Indirect Effects rCS,EX = g21 + g11b21 Direct effect: coefficient that links independent with dependent variable; e.g., g21 is direct effect of EX on CS Indirect effect: effect of one variable on another via one or more intervening variable(s), e.g., g11b21 Total indirect effect: sum of indirect effects between two variables Total effect: sum of direct and total indirect effects between two variables 3.12.2004 Messung & Analyse von Kundenzufriedenheit 26 Decomposition of Covariance syx s YX I (Y I (Y X) YI s IX X ): variable on path from X to Y YI: path coefficient of variable I to Y 3.12.2004 Messung & Analyse von Kundenzufriedenheit 27 First Law of Path Analysis Decomposition of covariance sxy between Y and X: s YX i(Y X ) Yis iX Assumptions: Exogenous (X) and endogenous variables (Y) have mean zero Errors or noises (z) have mean zero and equal variances across observations are uncorrelated across observations are uncorrelated with exogenous variables are uncorrelated across equations 3.12.2004 Messung & Analyse von Kundenzufriedenheit 28 Identification PQ = g11EX + z1 CS = b21PQ + g21EX + z2 Y1 = g11X + z1 Y2 = b21Y1 + g21X + z2 In matrix-notation: Y = BY + GX + z 2 0 0 g s1 0 11 2 B , G , (s EX ), 2 b 21 0 g 21 0 s2 Number of parameters: p=6 Model is identified, if all parameters can be expressed as functions of variances/covariances of observed variables 3.12.2004 Messung & Analyse von Kundenzufriedenheit 29 Identification, cont’d Y1 = g11X + z1 Y2 = b21Y1 + g21X + z2 s1X =g11 sX s2X = b21s1X + g21sX2 s21 = b21s12 + g21s1X sX 2 = sX 2 sy12 = g11s1X+s12 sy22 = b21s21 + g21s2X+s22 2 3.12.2004 p=6 first 3 equations allow unique solution for path coefficients, last three for variances of d and z Messung & Analyse von Kundenzufriedenheit 30 Condition for Identification Just-identified: all parameters can be uniquely derived from functions of variances/covariances Over-identified: at least one parameter is not uniquely determined Under-identified: insufficient number of variances/covariances Necessary, but not sufficient condition for identification: number of variances/covariances at least as large as number of parameters A general and operational rule for checking identification has not been found 3.12.2004 Messung & Analyse von Kundenzufriedenheit 31 Latent variables and Indicators Latent variables (LVs) or constructs or factors are unobservable, but We might find indicators or manifest variables (MVs) for the LVs that can be used as measures of the latent variable Indicators are imperfect measures of the latent variable 3.12.2004 Messung & Analyse von Kundenzufriedenheit 32 Indicators for “Expectation” d1 d2 d3 From: Swedish CSB Questionnaire, Banks: Private Customers E1 E2 E3 EX E1, E2, E3: „block“ of LVs for Expectation E1: When you became a customer of AB-Bank, you probably knew something about them. How would you grade your expectations on a scale of 1 (very low) to 10 (very high)? E2: Now think about the different services they offer, such as bank loans, rates, … Rate your expectations on a scale of 1 to 10? E3: Finally rate your overall expectations on a scale of 1 to 10? 3.12.2004 Messung & Analyse von Kundenzufriedenheit 33 Notation d1 d2 d3 X1 X2 l1 l2 l3 x X3 x: latent variable, factor Xi: indicators, manifest variables li: factor loadings di: measurement errors, noise 3.12.2004 X1=l1x+d1 X2=l2x+d2 X3=l3x+d3 “reflective” indicators Some properties: LV: unit variance noise di: has mean zero, variance si2, uncorrelated with other noises Messung & Analyse von Kundenzufriedenheit 34 Notation d1 d2 d3 X1 X2 l1 l2 l3 x X3 x: latent variable, factor Xi: indicators, manifest variables li: factor loadings di: measurement error, noise 3.12.2004 X1=l1x+d1 X2=l2x+d2 X3=l3x+d3 In matrix-notation: X = Lx + d with vectors X, L, and d e.g., X = (X1, X2, X3)‘ Messung & Analyse von Kundenzufriedenheit 35 CS-Model: Path Diagram d1 d2 d3 e1 e2 e3 3.12.2004 E1 E2 E3 EX Q2 Q3 g21 g11 Q1 PQ z2 C1 CS b21 e4 e5 C2 e6 C3 z1 Messung & Analyse von Kundenzufriedenheit 36 SEM-Model: Path Diagram d1 d2 d3 e1 e2 e3 3.12.2004 X1 X2 X3 x Y3 g21 h1 h2 e5 Y5 e6 Y6 b21 z1 e4 Y4 g11 Y1 Y2 z2 h = Bh + Gx + z X = Lxx+d, Y = Lyh+e Messung & Analyse von Kundenzufriedenheit 37 SEM-Model: Notation h = Bh + Gx + z Inner relations, inner model 2 s 0 0 0 g 11 2 1 B , G , (s EX ), 2 b 21 0 g 21 0 s2 Outer relations, measurement model X, d: 3-component vector X = Lxx+d, Y = Lyh+e Y, e: 6-component vector l11 l12 l13 0 0 0 Lx ( l11 l12 l13 ) , Ly , d , e 0 0 0 l11 l12 l13 3.12.2004 Messung & Analyse von Kundenzufriedenheit 38 Statistical Assumptions Error terms of inner model (z) have zero means constant variances across observations are uncorrelated across observations are uncorrelated with exogenous variables Error terms of measurement models (d, e) have zero means constant variances across observations are uncorrelated across observations are uncorrelated with latent variables and with each other Latent variables are standardized 3.12.2004 Messung & Analyse von Kundenzufriedenheit 39 Covariance Matrix of Manifest Variables Unrestricted covariance matrix (order: K = kx+ky) S = Var{(X’,Y’)’} Model-implied covariance matrix A1 A2 S( ) , (B, G, , , L x , L y , d , e ) A2 A3 A1 L x ( I - B) -1 (GG )[( I - B) -1 ]L y e A2 L y ( I - B) -1 G[L x ] A3 L x [L x ] d 3.12.2004 Messung & Analyse von Kundenzufriedenheit 40 Estimation of the Parameters Covariance fitting methods search for values of parameters so that the modelimplied covariance matrix fits the observed unrestricted covariance matrix of the MVs LISREL (LInear Structural RELations): Jöreskog (1973), Keesling (1972), Wiley (1973) Software LISREL by Jöreskog & Sörbom PLS techniques partition of in estimable subsets of parameters iterative optimizations provide successive approximations for LV scores and parameters Wold (1973, 1980) 3.12.2004 Messung & Analyse von Kundenzufriedenheit 41 Discrepancy Function The discrepancy or fitting function F(S;S) = F(S; S()) is a measure of the “distance” between the modelimplied covariance-matrix S() and the estimated unrestricted covariance-matrix S Properties of the discrepancy function: F(S;S) ≥ 0; F(S;S) = 0 if S=S 3.12.2004 Messung & Analyse von Kundenzufriedenheit 42 Covariance Fitting (LISREL) Estimates of the parameters are derived by F(S;S()) min Minimization of (K: number of indicators) F(S;S) = log|S| – log|S| + trace (SS-1) – K gives ML-estimates, if the manifest variables are independently, multivariate normally distributed Iterative Algorithm (Newton-Raphson type) Identification Choice of starting values is crucial Other choices of F result in estimation methods like OLS and GLS; ADF (asymptotically distribution free) 3.12.2004 Messung & Analyse von Kundenzufriedenheit 43 PLS Techniques Estimates factor scores for latent variables Estimates structural parameters (path coefficients, loading coefficients), based on estimated factor scores, using the principle of least squares Maximizes the predictive accuracy “Predictor specification”, viz. that E(h|x) equals the systematic part of the model, implies E(z|x)=0: the error term has (conditional) mean zero No distributional assumptions beyond those on 1st and 2nd order moments 3.12.2004 Messung & Analyse von Kundenzufriedenheit 44 The PLS-Algorithm Step 1: Estimation of factor scores 1. 2. 3. 4. Outer approximation Calculation of inner weights Inner approximation Calculation of outer weights Step 2: Estimation of path and loading coefficients by minimizing Var(z) and Var(d) Step 3: Estimation of location parameters (intercepts) Bo from h = Bo + Bh + Gx + z Lo from X = Lo + Lxx + d 3.12.2004 Messung & Analyse von Kundenzufriedenheit 45 Estimation of Factor Scores Factor hi: realizations Yin, n=1,…,N Yin(o): outer approximation of Yin Yin(i): inner approximation of Yin Indicator Yij: observations yijn; j=1,…,Ji; n=1,…,N 1. Outer approximation: Yin(o)=Sjwijyijn s.t. Var(Yi(o))=1 2. Inner weights: vih=sign(rih), if hi and hh adjacent; otherwise vih=0; rih=corr(hi,hh) (“centroid weighting”) 3. Inner approximation: Yin(i)=ShvihYhn(o) s.t. Var(Yi(i))=1 4. Outer weights: wij=corr(Yij,Yi(i)) Start: choose arbitrary values for wij Repeat 1. through 4. until outer weights converge 3.12.2004 Messung & Analyse von Kundenzufriedenheit 46 Example d1 d2 d3 e1 e2 e3 3.12.2004 E1 E2 E3 EX Q2 Q3 g21() g11() Q1 PQ z2 C1 CS b21() e4 e5 C2 e6 C3 z1 Messung & Analyse von Kundenzufriedenheit 47 Example, cont’d Starting values wEX,1,…,wEX,3,wPQ,1,…,wPQ,3,wCS,1,…,wCS,3 Outer approximation: EXn(o) = SjwEX,jEjn; similar PQn(o), CSn(o); standardized Inner approximation: EXn(i) = + PQn(o) + CSn(o) PQn(i) = + EXn(o) + CSn(o) CSn(i) = + EXn(o) + PQn(o) standardized Outer weights: wEX,j = corr(Ej,EX(i)), j=1,…,3; similar wPQ,j, wCS,j 3.12.2004 Messung & Analyse von Kundenzufriedenheit 48 Choice of Inner Weights Centroid weighting scheme: Yin(i)=ShvihYhn(o) vij=sign(rih), if hi and hh adjacent, vij=0 otherwise with rih=corr(hi,hh); these weights are obtained if vih are chosen to be +1 or -1 and Var(Yi(i)) is maximized Weighting schemes: hh predecessor centroid hh successor sign(rih) sign(rih) factor, PC rih rih path bih rih bih: coefficient in regression of hi on hh 3.12.2004 Messung & Analyse von Kundenzufriedenheit 49 Measurement Model: Examples Latent variables from Swedish CSB Model 1. Expectation 2. E1: new customer feelings E2: special products/services expectations E3: overall expectation Perceived Quality Q1: Q2: Q3: Q4: Q5: 3.12.2004 range of products/services quality of service clarity of information on products/services opening hours and appearance of location etc. Messung & Analyse von Kundenzufriedenheit 50 Measurement Models Reflective model: each indicator is reflecting the latent variable (example 1) Yij = lijhi + eij Yij is called a reflective or effect indicator (of hi) Formative model: (example 2) hi = py'Yi + di py is a vector of ki weights; Yij are called formative or cause indicators Hybrid or MIMIC model (for “multiple indicators and multiple causes”) Choice between formative and reflective depends on the substantive theory Formative models often used for exogenous, reflective and MIMIC models for endogenous variables 3.12.2004 Messung & Analyse von Kundenzufriedenheit 51 Estimation of Outer Weights “Mode A” estimation of Yi(o): reflective measurement model weight wij is coefficient from simple regression of Yi(i) on Yij: wij = corr(Yij,Yi(i)) “Mode B” estimation of Yi(o): formative measurement model weight wij is coefficient of Yij from multiple regression of Yi(i) on Yij, j=1,…,Ji multicollinearity?! MIMIC model 3.12.2004 Messung & Analyse von Kundenzufriedenheit 52 Properties of Estimators A general proof for convergence of the PLS-algorithm does not exists; practitioners experience no problems Factor scores are inconsistent but “consistent at large”: consistency is achieved with increasing sample size and block size Loading coefficients are inconsistent and seem to be overestimated Path coefficients are inconsistent and seem to be underestimated 3.12.2004 Messung & Analyse von Kundenzufriedenheit 53 ACSI Model: Results Perceived Quality 0,90 0,73 0,95 0,53 Expectations 3.12.2004 -0,38 -0,29 0,78 0,47 (-0,15) 0,12 -0,24 (0,06) Value Customer Satisfaction 0,57 0,35 0,40 0,35 Voice 0,17 (0,06) Loyalty EQS-estimates PLS-estimates Messung & Analyse von Kundenzufriedenheit 54 Evaluation of SEM-Models Depends on estimation method Covariance-fitting methods: distributional assumptions, optimal parameter estimates, factor indeterminacy PLS path modeling: non-parametric, optimal prediction accuracy, LV scores Step 1: Inspection of estimation results (R2, parameter estimates, standard errors, LV scores, residuals, etc.) Step 2: Assessment of fit Covariance-fitting methods: global measures PLS path modeling: partial fitting measures 3.12.2004 Messung & Analyse von Kundenzufriedenheit 55 Inspection of Results Covariance-fitting methods: global optimization Model parameters and their standard errors; do they confirm theory? Correlation residuals: sij-sij() Graphical methods PLS techniques: iterative optimization of outer models and inner model Model parameters Resampling procedures like blindfolding or jackknifing give standard errors of model parameters LV scores Graphical methods 3.12.2004 Messung & Analyse von Kundenzufriedenheit 56 Fit Indices Covariance-fitting methods: covariance fit measures such as Chi-square statistics Goodness of Fit Index (GFI), AGFI Normed Fit Index (NFI), NNFI, CFI Etc. Basis is the discrepancy function PLS path modeling: prediction-based measures Communality Redundancy Stone-Geisser’s Q2 3.12.2004 Messung & Analyse von Kundenzufriedenheit 57 Chi-square Statistic Test of H0: S = S() against non-specified alternative Test-statistic X2=(N-1)F(S;S(̂)) If model is just identified (c=p): X2=0 [c=K(K+1)/2, p: number of parameters in ] Under usual regularity conditions (normal distribution, ML-estimation), X2 is asymptotically 2(c-p)-distributed Non-significant X2 indicate: the over-identified model does not differ from a just-identified version Problem: X2 increases with increasing N Some prefer X2/(c-p) to X2 (has reduced sensitivity to sample size); rule of thumb: X2/(c-p) < 3 is acceptable 3.12.2004 Messung & Analyse von Kundenzufriedenheit 58 Goodness of Fit Indices Goodness of Fit Index (Jöreskog & Sörbom): F [ S , S(ˆ)] GFI 1 - F [ S , S(O)] Portion of observed covariances explained by the model-implied covariances “How much better fits the model as compared to no model at all” Ranges from 0 (poor fit) to 1 (perfect fit) Rule of thumb: GFI > 0.9 AGFI penalizes model complexity: c F [ S , S(ˆ)] AGFI 1 - c p F [ S , S(O)] 3.12.2004 Messung & Analyse von Kundenzufriedenheit 59 Other Fit Indices Normed Fit Index, NFI (Bentler & Bonett) Comparative Fit Index, CFI (Bentler) Less depending of sample size than NFI Non-Normed Fit Index, NNFI (Bentler & Bonett) Similar to GFI, but compares with a baseline model, typically the independence model (indicators are uncorrelated) Ranges from 0 (poor fit) to 1 (perfect fit) Rule of thumb: NFI > 0.9 Also known as Tucker-Lewis Index Adjusted for model complexity Root mean squared error of approximation, RMSEA (Steiger): RMSEA F[S , S(ˆ)]/(c - p) 3.12.2004 Messung & Analyse von Kundenzufriedenheit 60 Assessment of PLS Results Not a single but many optimization steps; not a global measure but many measures of various aspects of results Indices for assessing the predictive relevance Portions of explained variance (R2) Communality, redundancy, etc. Stone-Geisser’s Q2 Reliability indices NFI, assuming normality of indicators Allows comparisons with covariance-fitting results 3.12.2004 Messung & Analyse von Kundenzufriedenheit 61 Some Indices Assessment of diagonal fit (proportion of explained variances) SMC (squared multiple correlation coefficient) R2: (average) proportion of the variance of LVs that is explained by other LVs; concerns the inner model Communality H2: (average) proportion of the variance of indicators that is explained by the LVs directly connected to it; concerns the outer model Redundancy F2: (average) proportion of the variance of indicators that is explained by predictor LVs of its own LV r2: proportion of explained variance of indicators 3.12.2004 Messung & Analyse von Kundenzufriedenheit 62 Some Indices, cont’d Assessment of non-diagonal fit Explained indicator covariances rs = 1- c/s with c = rms(C), s = rms(S); C: estimate of Cov(e) Explained latent variable correlation rr = 1- q/r with q = rms(Q), r = rms(Cov(Y)); Q: estimate of Cov(z) reY = rms (Cov(e,Y)), e: outer residuals reu = rms (Cov(e,u)), u: inner residuals rms(A) = (SiSj aij2)1/2: root mean squared covariances (diagonal elements of symmetric A excluded from summation) 3.12.2004 Messung & Analyse von Kundenzufriedenheit 63 Stone-Geisser’s Q2 Similar to R2 E Q 1O E: sum of squared prediction errors; O: sum of squared deviations from mean Prediction errors from resampling (blindfolding, jackknifing) E.g., communality of Yij, an indicator of hi ˆ Y )]2 [ y ( l ijn ij in Qijc2 1 - n n [ yijn - yij ]2 2 3.12.2004 Messung & Analyse von Kundenzufriedenheit 64 Lohmöller’s Advice Check fit of outer model Low unexplained portion of indicator variances and covariances High communalities in reflective blocks, low residual covariances Residual covariances between blocks close to zero Covariances between outer residuals and latent variables close to zero Check fit of inner model Low unexplained portion of latent variable indicator variances and covariances Check fit of total model High redundancy coefficient Low covariances of inner and outer residuals 3.12.2004 Messung & Analyse von Kundenzufriedenheit 65 ACSI Model: Results Perceived Quality 0,90 0,73 0,95 0,53 Expectations 3.12.2004 -0,38 -0,29 0,78 0,47 (-0,15) 0,12 -0,24 (0,06) Value Customer Satisfaction 0,57 0,35 0,40 0,35 Voice 0,17 (0,06) Loyalty EQS-estimates PLS-estimates Messung & Analyse von Kundenzufriedenheit 66 Diagnostics: EQS 3.12.2004 ACSI ACSIe 2 247.5 378.7 df 81 NNFI 0.898 0.930 RMSEA 0.079 0.060 173 Messung & Analyse von Kundenzufriedenheit 67 Diagnostics: PLS (centroid weighting) ACSI 3.12.2004 ACSI e Hui Schenk R2 0.29 0.35 0.43 0.40 Q2 0.36 0.41 0.58 0.49 rr 0.47 0.55 0.58 0.59 H2 0.71 0.64 0.64 0.64 F2 0.22 0.24 0.30 0.26 r2 0.63 0.63 0.57 0.60 reY 0.26 0.24 0.19 0.09 reu 0.19 0.17 0.16 0.08 Messung & Analyse von Kundenzufriedenheit 68 Practice of SEM Analysis Theoretical basis Data Scaling: metric or nominal (in LISREL not standard) Sample-size: a good choice is 10p (p: number of parameters); <5p cases might result in unstable estimates; large number of cases will result in large values of X2 Reflective indicators are assumed to be uni-dimensional; it is recommended to use principal axis extraction, Cronbach’s alpha and similar to confirm the suitability of data Model Identification must be checked for covariance fitting methods Indicators for LV can be formative or reflective; formative indicators not supported in LISREL 3.12.2004 Messung & Analyse von Kundenzufriedenheit 69 Practice of SEM Analalysis cont’d Model LISREL allows for more general covariance structures e.g., correlation of measurement errors Estimation Repeat estimation with varying starting values Diagnostic checks Use graphical tools like plots of residuals etc. Check each measurement model Check each structural equation Lohmöller’s advice Model trimming Stepwise model building (Hui, 1982; Schenk, 2001) 3.12.2004 Messung & Analyse von Kundenzufriedenheit 70 LISREL vs PLS Models PLS assumes recursive inner structure PLS allows for higher complexity w.r.t. B, G, and L; LISREL w.r.t. and Estimation method Distributional assumptions in PLS not needed Formative measurement model in PLS Factor scores in PLS PLS: biased estimates, consistency at large LISREL: ML-theory In PLS: diagnostics much richer Empirical facts LISREL needs in general larger samples LISREL needs more computation 3.12.2004 Messung & Analyse von Kundenzufriedenheit 71 The Extended Model Perceived Quality 0,87 0,73 0,85 0,53 Expectations 3.12.2004 (0,20) 0,33 0,58 0,37 (-0,14) (0,06) Emotional Factor 0,55 0,36 (-0,14) (-0,01) Value 0,31 0,35 Customer Satisfaction 0,48 0,34 0,41 0,34 Loyalty EQS-estimates PLS-estimates Messung & Analyse von Kundenzufriedenheit 72 Diagnostics: EQS 3.12.2004 ACSI ACSI 2 247.5 378.7 df 81 NNFI 0.898 0.930 RMSEA 0.079 0.060 e 173 Messung & Analyse von Kundenzufriedenheit 73 Diagnostics: PLS (centroid weighting) ACSI 3.12.2004 ACSI e Hui Schenk R2 0.29 0.35 0.43 0.40 Q2 0.36 0.41 0.58 0.49 rr 0.47 0.55 0.58 0.59 H2 0.71 0.64 0.64 0.64 F2 0.22 0.24 0.30 0.26 r2 0.63 0.63 0.57 0.60 reY 0.26 0.24 0.19 0.09 reu 0.19 0.17 0.16 0.08 Messung & Analyse von Kundenzufriedenheit 74 Model Building: Hui’s Approach Perceived Quality 0,43 Emotional Factor 0,31 -0,18 0,10 0,33 -0,18 3.12.2004 0,35 0,36 0,42 Expectations 0,61 0,17 0,63 Value 0,12 Customer Satisfaction 0,21 0,23 Messung & Analyse von Kundenzufriedenheit Loyalty 75 Model Building: Schenk’s Approach 0,32 Perceived Quality Emotional Factor 0,35 0,31 0,32 0,73 0,34 0,60 Expectations 3.12.2004 Customer Satisfaction Value Messung & Analyse von Kundenzufriedenheit 76 The end 3.12.2004 Messung & Analyse von Kundenzufriedenheit 77 Data-driven Specification No solid a priori knowledge about relations among variables Stepwise regression Search of the “best” model Forward selection Backward elimination Problem: omitted variable bias General to specific modeling 3.12.2004 Messung & Analyse von Kundenzufriedenheit 78 Stepwise SE Model Building Hui (1982): models with interdependent inner relations Schenk (2001): guaranties causal structure, i.e., triangular matrix B of path coefficients in the inner model η=Bη+ζ 3.12.2004 Messung & Analyse von Kundenzufriedenheit 79 Stepwise SE Model Building Hui’s algorithm Stage 1 1. Calculate case values Yij for LVs ηi as principal component of corresponding block, calculate R = Corr(Y) 2. Choose for each endogenous LV the one with highest correlation to form a simple regression 3. Repeat until a stable model is reached a. PLS-estimate the model, calculate case values, and recalculate R b. Drop from each equation LVs with t-value |t|<1,65 c. Add in each equation the LV with highest partial correlation with dependent LV 3.12.2004 Messung & Analyse von Kundenzufriedenheit 80 Stepwise SE Model Building Hui’s algorithm, cont’d Stage 2 1. Use rank condition for checking identifiability of each equation 2. Use 2SLS for estimating the path coefficients in each equation 3.12.2004 Messung & Analyse von Kundenzufriedenheit 81 Hui’s vs. Schenk’s Algorithm Hui’s algorithm is not restricted to a causal structure; allows cycles and an arbitrary structure of matrix B Schenk’s algorithm 3.12.2004 uses an iterative procedure similar to that used by Hui makes use of a priori information about the structure of the causal chain connecting the latent variables latent variables are to be sorted Messung & Analyse von Kundenzufriedenheit 82 Stepwise SE Model Building Schenk’s algorithm 1. Calculate case values Yij for LVs ηi as principal component of corresponding block, calculate R = Corr(Y) 2. Choose pair of LVs with highest correlation 3. Repeat until a stable model is reached a. PLS-estimate the model, calculate case values, and recalculate R b. Drop LVs with non-significant t-value c. Add LV with highest correlation with already included LVs 3.12.2004 Messung & Analyse von Kundenzufriedenheit 83 Data, special CS dimensions Staff 2 availability1 (PERS), politeness1 (FREU) Outlet 3 make-up1 (GEST), presentation of merchandise1 (PRAE), cleanliness1 (SAUB) Range 2 freshness and quality (QUAL), richness (VIEL) Customerorientation 7 access to outlet (ERRE), shopping hours (OEFF), queuing time for checkout1 (WART), paying modes1 (ZAHL), price information1 (PRAU), sales (SOND), availability of sales (VERF) 1 Dimension of “Emotional Factor” 3.12.2004 Messung & Analyse von Kundenzufriedenheit 84 References C. Fornell (1992), “A National Customer Satisfaction Barometer: The Swedish Experience”. Journal of Marketing, (56), 6-21. C. Fornell and Jaesung Cha (1994), “Partial Least Squares”, pp. 52-78 in R.P. Bagozzi (ed.), Advanced Methods of Marketing Research. Blackwell. J.B. Lohmöller (1989), Latent variable path modeling with partial least squares. Physica-Verlag. H. Wold (1982), “Soft modeling. The basic design and some extensions”, in: Vol.2 of Jöreskog-Wold (eds.), Systems under Indirect Observation. North-Holland. H. Wold (1985), “Partial Least Squares”, pp. 581-591 in S. Kotz, N.L. Johnson (eds.), Encyclopedia of Statistical Sciences, Vol. 6. Wiley. 3.12.2004 Messung & Analyse von Kundenzufriedenheit 85