Using the SmartPLS Software “Structural Model Assessment” Joe F. Hair, Jr. Founder & Senior Scholar All rights reserved ©. Cannot be reproduced or distributed without express written permission from Sage, Prentice-Hall, SmartPLS, and session presenters. 1 Step 3: Assess the Level of R2 The R² values of the endogenous latent variables are available in the PLS Algorithm default report (PLS → Quality Criteria → Overview), as shown below. The R2 values of COMP (0.6309), CUSL (0.5620), and LIKE (0.5576) can be considered moderate. In contrast, the R² value of CUSA (0.2919) is rather weak. All rights reserved ©. Cannot be reproduced or distributed without express written permission from Sage, Prentice-Hall, SmartPLS, and session presenters. 2 2 Note: calculation of Q2 value is shown in a later slide. 3 Step 4: Assessing Effect Size – ƒ 2 The ƒ² effect size is a measure of the impact of a specific predictor construct on an endogenous construct. In addition to evaluating the size of the R² values of all endogenous constructs, the ƒ² effect size can be calculated (it is not available from the SmartPLS software output). The ƒ² effect size measures the change in the R² value when a specified exogenous construct is omitted from the model. It is used to evaluate whether the omitted predictor construct has a substantive impact on the R² values of the endogenous construct(s). The formula for calculating the ƒ² effect size is shown on the next slide. Guidelines for assessing ƒ2 values for the exogenous latent constructs in predicting the endogenous constructs are: Value 0.02 0.15 0.35 = = = Effect Size small medium large (Cohen, 1988) All rights reserved ©. Cannot be reproduced or distributed without express written permission from Sage, Prentice-Hall, SmartPLS, and session presenters. 4 Calculating the Effect Size – ƒ 2 The f 2 effect size can be calculated as shown below. In the formula R2included and R2excluded are the R² values of the endogenous latent variable when a selected exogenous latent variable is included or excluded from the model. The change in the R² values is calculated by estimating the PLS path model twice. It is estimated the first time with the exogenous latent variable included (yielding R2included) and the second time with the exogenous latent variable excluded (yielding R2excluded). The results of calculating the f 2 for the reputation model example endogenous variables are shown in a later slide. All rights reserved ©. Cannot be reproduced or distributed without express written permission from Sage, Prentice-Hall, SmartPLS, and session presenters. 5 Example: Calculation of f2 Effect Size As indicated in the previous slide, to compute the f2 value of a selected endogenous latent construct, we need the R2 included and R2excluded values. The R2included results from the overall model estimation were previously shown (Exhibit 6.15). The R2excluded value is obtained from a model re-estimation after deleting a specific predecessor of that endogenous latent variable. For example, the endogenous latent variable CUSL has an original R2 value of 0.562 (R2included). If CUSA is deleted from the path model and the model is re-estimated, the R2 of CUSL has a value of only 0.385 (R2excluded). These two values are the inputs for computing the f2 effect size of CUSA on CUSL. The formula is shown below: 6 Example: Calculation of f2 for other endogenous constructs To get the f2 values you need to run the full model first and determine the R2 for the endogenous construct you want to evaluate. Next, eliminate one path pointing at the construct you are looking at (simply right click on a construct and delete the predictor construct), and re-run the model. The R2 with the construct/path will be lower since the predictor construct was removed. Now enter the R2 included for the selected construct (based on the full model) and the R2 excluded (based on the reduced model where one path/construct has been deleted) into the formula on the previous slide. The same procedure is followed for the q2 but instead of entering the R2 (excluded and included), you use blindfolding to get the Q2 values for the full model (included) and the reduced model (construct/path deleted). All rights reserved ©. Cannot be reproduced or distributed without express written permission from Prentice-Hall, McGraw-Hill, SmartPLS, and session presenters. 7 Summary of Results – Path Coefficients, f2 and q2 COMP CUSL Path Coefficient f2 effect size q2 effect size 0.0057 0.000 0.101 CUSA 0.5050 0.404 0.229 LIKE 0.3440 0.139 0.081 ATTR Path Coefficient f2 effect size q2 effect size 0.0861 0.011 0.028 COMP CSOR 0.0589 0.005 0.004 PERF 0.2955 0.076 0.042 QUAL 0.4297 0.144 0.062 Example interpretation of f2: Look under the f2 column for CUSL. Note the 0.404 is the f2 effect size for the predictive value of CUSA on CUSL. The 0.404 indicates that CUSA has a large effect in producing the R2 for CUSL. In contrast, the 0.139 is the f2 effect size for the predictive value of LIKE on CUSL. The 0.139 indicates that LIKE has close to a medium effect in 8 producing the R2 for CUSL. Reputation Model Results – Path Coefficients – Guidelines for assessing ƒ2 values: Value Effect Size 0.02 = small 0.15 = medium All rights reserved ©. Cannot be reproduced or distributed without express written permission 9 0.35 = large from Sage, Prentice-Hall, SmartPLS, and session presenters. Summary of Results – Path Coefficients, f2 and q2 CUSA Path Coefficient f2 effect size q2 effect size 0.1671 0.029 0.019 0.1784 0.036 0.029 PERF 0.1170 0.011 0.007 QUAL 0.3800 0.095 0.053 Path Coefficient f2 effect size LIKE q2 effect size ATTR COMP 0.1455 0.018 0.001 CSOR CUSA LIKE 0.4357 0.161 0.140 10 Step 5: Blindfolding and Predictive Relevance – Q 2 In addition to evaluating the magnitude of the R² values as a criterion of predictive accuracy, researchers should also examine the Q² value – which is an indicator of the model’s predictive relevance. The Q² measure applies a sample re-use technique that omits part of the data matrix and uses the model estimates to predict the omitted part. Specifically, when a PLS-SEM model exhibits predictive relevance, it accurately predicts the data points of the indicators in reflective measurement models of multi-item as well as single-item endogenous constructs (the procedure does not apply to formative endogenous constructs). For SEM models, Q² values larger than zero for a specific reflective endogenous latent variable indicate the path model’s predictive relevance for a particular construct. Q² values of zero or below indicate a lack of predictive relevance. As a relative measure of predictive relevance, values of 0.02, 0.15, and 0.35 indicate that an exogenous construct has a small, medium, or large predictive relevance for a selected endogenous construct. All rights reserved ©. Cannot be reproduced or distributed without express written permission from Sage, Prentice-Hall, SmartPLS, and session presenters. 11 Blindfolding and Predictive Relevance – Q 2 The Q ² value can be calculated by using two different approaches. The cross-validated redundancy approach uses the path model estimates of both the structural model (scores of the antecedent constructs) and the measurement model (target endogenous constructs). An alternative method is the cross-validated communality approach. This method uses only the construct scores estimated for the target endogenous construct (without including the structural model information) to predict the omitted data points. We recommend using the cross-validated redundancy as a measure of Q2 since it includes the key element of the path model, the structural model, to predict eliminated data points. When you run the blindfolding option for cross-validated redundancy, all constructs in your SEM model are shown (see next slide). You can select multiple latent variables in the dialog box, but you need to run this routine for one reflective target construct at a time – one after the other until all have been tested. Finally, this option is only used with endogenous constructs that are measured reflectively. All rights reserved ©. Cannot be reproduced or distributed without express written permission from Sage, Prentice-Hall, SmartPLS, and session presenters. 12 SmartPLS Predictive Relevance – Blindfolding Redundancy vs. Communality? Cross-validated redundancy Step 1: The scores of the endogenous LV(s) are estimated using the scores of the exogenous LVs LV1 MV 1 LV3 MV 3 LV2 LV1 Step 2: Newly estimated LV scores are used to estimate the missing MV data MV 1 LV3 MV 2 MV 3 LV2 Cross-validated communality MV 2 Only step 2. . Running Blindfolding to obtain Q2 for Endogenous Construct CUSL Note that only CUSL is checked. All rights reserved ©. Cannot be reproduced or distributed without express written permission from Sage, Prentice-Hall, SmartPLS, and session presenters. 14 Blindfolding Results for Endogenous Construct CUSL . Q2 value for CUSL Full Base Model The result column is at the top right corner (1 – SSE/SSO). For our path model the predictive relevance Q2 of CUSL has a value of 0.4178, which indicates the model has large predictive relevance for this construct. When blindfolding is run for all endogenous latent constructs in the model they all have Q2 values considerably All rights reserved ©. Cannot be reproduced or distributed without express written permission above as shownSmartPLS, on theand next slide. from zero, Sage, Prentice-Hall, session presenters. 15 Reputation Model – R2 and Q2 Measures The table above shows that all Q2 values are considerably above zero, thus providing support for the reputation model’s predictive relevance for the four endogenous constructs. Computation of q 2 The final assessment addresses the calculation of the q² effect sizes. The calculation of q2 for the CUSL construct of the reputation model is shown below. To compute the q² value of a selected endogenous latent variable, you need the Q2included and Q2excluded values. The Q2included results for all endogenous constructs from the overall model estimation are available from a previous slide. The Q2excluded value is obtained from a model re-estimation after deleting a specific predecessor of that endogenous latent variable. For example, the endogenous latent variable CUSL has an original Q² value of 0.418 (Q2included). If CUSA is deleted from the path model and the model is re-estimated, the Q² of CUSL has a value of only 0.285 (Q2excluded ). These two values are the inputs for computing the q² effect size of CUSA. on CUSL, as shown below: All rights reserved ©. Cannot be reproduced or distributed without express written permission from Sage, Prentice-Hall, SmartPLS, and session presenters. 17 PLS Algorithm Results Base Model with CUSA Removed . All rights reserved ©. Cannot be reproduced or distributed without express written permission from Sage, Prentice-Hall, SmartPLS, and session presenters. 18 Blindfolding Results for CUSL with full Base Model . Q2 value for CUSL with full model All rights reserved ©. Cannot be reproduced or distributed without express written permission from Sage, Prentice-Hall, SmartPLS, and session presenters. 19 Blindfolding Results for CUSL Reduced Model with CUSA Removed Q2 value for CUSL with CUSA Removed . All rights reserved ©. Cannot be reproduced or distributed without express written permission from Sage, Prentice-Hall, SmartPLS, and session presenters. 20 Summary of Results – Path Coefficients, f2 and q2 COMP CUSL Path Coefficient f2 effect size q2 effect size 0.0057 0.000 0.101 CUSA 0.5050 0.404 0.229 LIKE 0.3440 0.139 0.081 ATTR Path Coefficient f2 effect size q2 effect size 0.0861 0.011 0.028 COMP CSOR 0.0589 0.005 0.004 PERF 0.2955 0.076 0.042 QUAL 0.4297 0.144 0.062 Example interpretation of q2: Look under the q2 column for CUSL. Note the 0.229 is the q2 effect size for the predictive relevance of CUSA on CUSL. The 0.229 indicates that CUSA has a medium effect in producing the Q2 (predictive relevance) for CUSL. In contrast, the 0.081 is the q2 effect size for the predictive relevance of LIKE on CUSL. The 0.081 indicates that LIKE has a 21 small effect in producing the Q2 for CUSL. Reputation Model Results – Path Coefficients – Guidelines for assessing q2 values: Value Effect Size 0.02 = small 0.15 = medium All rights reserved ©. Cannot be reproduced or distributed without express written permission 22 0.35 = large from Sage, Prentice-Hall, SmartPLS, and session presenters. Summary of Results – Path Coefficients, f2 and q2 CUSA Path Coefficient f2 effect size LIKE q2 effect size ATTR COMP 0.1455 0.018 Path Coefficient f2 effect size q2 effect size 0.1671 0.029 0.019 0.1784 0.036 0.029 0.1170 0.011 0.007 0.3800 0.095 0.053 0.001 CSOR CUSA LIKE PERF QUAL 0.4357 0.161 0.140 Guidelines for assessing q2 values associated with predictive relevance (Q2): Value Effect Size 0.02 = small 0.15 = medium 0.35 = large 23 . . All rights reserved ©. Cannot be reproduced or distributed without express written permission from Sage, Prentice-Hall, SmartPLS, and session presenters. 25