본문 바로가기
Choi's Research/1.Progress in Research

PLS-SEM Guidelines and Compliances Summary

by Jeonghwan (Jerry) Choi 2017. 7. 15.

PLS-SEM Guideline and Compliance Summary


A study of Diffusion of Legal Software Use in a Global Campus: Action Research at a Global Campus in China. 


by Jeonghwan (Jerry) Choi, Aug. 2017


Partial Least Square Structural Equation Modeling (PLS-SEM)

PLS-SEM method gets a special attention from business and social science researchers for its robustness in psychometric model analysis. This critical advantage makes the PLS-SEM as an alternative technique for SEM and it has become a key research method in recent years (J.F. Hair & Hult, 2016). For example, PLS-SEM gives a more robust structural equation model convergence over CB-SEM in many situations especially when a research model has many indicators, paths, and relationships among key variables and constructs (Chin, 2010; J.F. Hair & Hult, 2016; Henseler et al., 2014). For this benefit of PLS-SEM, it gets popular in management disciplines such as marketing management, strategic management, human resources management, and particularly management information system field of studies (J.F. Hair & Hult, 2016; J.F. Hair, Ringle, & Sarstedt, 2011; Williams et al., 2015).

Considering the rules of thumb for selecting CB-SEM or PLS-SEM (J.F. Hair et al., 2011), we judged that our study was appropriate for applying the PLS-SEM technique because this study’s main goal was to explore new knowledge and to extend an existing structural theory of the complex UTAUT. Prior to conducting the PLS-SEM, we summarized prerequisite or necessary conditions and recommendations through synthesizing a few foundational literature  (Chin, 2010; J.F. Hair & Hult, 2016; Latan & Ramli, 2013; Richter, Sinkovics, Ringle, & Schlaegel, 2016). By using our data, we validated these seven categories of PLS-SEM guidelines:

·      Data and sampling characteristics

·      PLS-SEM algorithm

·      Outer model evaluation: reflective (mode A)

·      Out model evaluation: formative (mode B)

·      Inner model evaluation: recursive model

·      Model fit

·      Multi-group analysis

Results of validation of PLS-SEM guideline and compliance are described in the following Table.

  

Overall, our data and research model showed good compliances to the proposed PLS-SEM guidelines. It should be noted that we use the Smart PLS [ SmartPLS (v. 3.2.6).  Ringle, Wende, and Becker, 2015. "SmartPLS 3." Boenningstedt: SmartPLS GmbH, http://www.smartpls.com] as the PLS-SEM analysis tool.




Table

PLS-SEM guideline and compliance 

Characteristics

Guideline

Recommendation

Compliance

Data and sampling characteristics

Sample size

Ten times rule: the minimum sample size should be equal to the larger of 10x largest number of 1) formative indicator and structural paths directed at a particular latent construct (Wong, 2013)

 

Alternatively, Cohen’s sample size recommendation of statistical power and effect sizes takes into accounts. Cohen’s sample size recommendation (In this study, 5% significance level, 0.10 minimum R2, and maximum number of arrows pointing at a construct (5) = 205

·  10x formative indicator (N/A)

·  10x structural paths directed a a latent construct (10 x 5 = 50)

 

·  Cohen’s sample size recommendation = 205

Samples size 

n = 215 > 205

Holdout

30% of original sample (Hair Jr & Hult, 2016)

> 30%

215/218 = 98.6%

Missing data

Less than 5% of missing or screening out data

< 5%

3 / 218 data were screened out = 1.38%

Distribution

Robust when applied to highly skewed data, but skewness and − kurtosis should be reported (Richter, Sinkovics, Ringle, & Schlaegel, 2016)

 

 

PLS-SEM Algorithm

Weighting scheme

In general, the path weighting scheme is strongly recommended because it provides the highest R2 value for endogenous latent variable (Vinzi, Chin, Henseler, & Wang, 2010)

path weighting

path weighting

Data metric

The standardized value setting

Mean 0, Var 1

Mean 0, Var 1

Total maximum iteration

The standard maximum iteration is 300

300

300

Abort criterion

The recommended number is 1.0E-5

1.0E-5

1.0E-7

Starting value

Initial outer weight can be set as 1.0

1.0

1.0

Algorithm to handle missing data

Missing value treatment options are mean replacement, EM (expectation-maximization algorithm), and nearest neighbor (Hair, Ringle, & Sarstedt, 2013)

Mean replacement

Mean replacement

Bootstrap subsample size

The number of bootstrap samples should be high but must be at least equal to the number of valid observations. As a rule, 5,000 bootstrap samples is recommended (Hair Jr & Hult, 2016, p. 132)

2000 - 4000

5000

Bootstrap sign change

No sigh change option is recommendable because it results in the most conservative outcome (Hair Jr & Hult, 2016, p. 135)

No sign change

No sign change

Significance level 

Generally, 5% significance level is widely used in social science field of study

5%, Two-tailed

5%, Two-tailed

Predictive relevance

In the process of blindfolding, omission distance (D) can be set between 5 and 10 (Hair Jr & Hult, 2016, p. 179)

£ D £ 10

D = 7

Software feature

SmartPLS (v. 3.2.6).  Ringle, C. M., Wende, S., and Becker, J.-M. 2015. "SmartPLS 3." Boenningstedt: SmartPLS GmbH, http://www.smartpls.com

-

Outer model evaluation: Reflective (mode A)

Indicator reliability

Recommended > 0.6 for exploratory research and > 0.7 for confirmatory research (Chin, 2010)

> 0.7

All indicators (factor loadings) are higher than 0.7 [0.737 ~ 0.939] 

Internal consistency reliability

The cut-off value for composite reliability is > 0.6 for exploratory research and > 0.7 for confirmatory research. The Cronbach’s alpha is not suggested for distinguishing 

> 0.7

All composite reliabilities are higher than 0.7 [0.904 ~ 0.952]

Convergent validity

The Average Variance Extracted (AVE) is > 0.5

AVE > 0.5

All AVE is higher than 0.5 [0.662 ~ 0.869]

Discriminant validity

Fornell and Larcker (1981) criterion: Each construct’s AVE should be higher than its squared correlation with any other construct (Fornell & Larcker, 1981)

Square root AVE > Correlation

All square root AVE is larger than any other correlations with other constructs

 

Cross-loading: Each indicator should load highest on the construct it is intended to measure (Chin, 2010)

Highest loading on the construct

Each indicator loaded highest on the intended construct

 

Heterotrait-Monotrait Ratio (HTMT) should under 0.85 for each outer model indicators: (Henseler, Hubona, & Ray, 2016)

HTMT ratio < 0.85

All HTMT ratio is under 0.85 [0.358 ~ 0.824]  except ‘Effort expectation and Facilitation’ [0.852] 

Item removal

If some item has been dropped to achieve a model fit, give additional information

No removed item

Outer model evaluation: Formative

 

According to Confirmatory Tetrad Analysis (CTA), all constructs in the model are not formative constructs in 1% level of significance (Hair Jr & Hult, 2016, pp. 46-47)

 

 

Collinearity

The cut-off value for VIF should be smaller than 0.5. A stabilized estimation is suggested as ranging 2.5 ~ 3.3 (Hair et al., 2013)

VIF < 0.5

Most of outer VIF values are under 0.5 [1.000 ~ 4.014]. However, the VIF values of UP1 (6.070) and UP2 (6.425) are higher than the criterion.    

Construct removal

If a construct has been dropped due to collinearity, the problem should be reported. 

 

No removed construct

Inner model evaluation: Recursive model

Path estimates

Reporting

1) Path coefficient

2) significance and confidence interval from bootstrapping

Bootstrapping is applied for the significant of the path coefficient with two-tails of 5% = 1.96

Bootstrapping is applied for the significant of the path coefficient with two-tails of 5% = 1.96  

R2  

(Adjusted R2)

R2 acceptable level is context-dependent. (Hair Jr & Hult, 2016; Latan & Ramli, 2013)

0.25: Weak

0.50: Moderate

0.75: Strong

R(Adj- R2) of Behavioral intention: 0.584 (0.578)

 

Rof Use Behavior 0.180 (0.172)

Effect size f2

Cohen’s statistical power analysis of effect size (Cohen, 1992)

0.02: Weak

0.15: Moderate

0.35: Strong

PE: 0.169

EE: 0.053

SI: 0.044

BI: 0.040

FA: 0.016

BI: 0.040

Predictive relevance

The cross validated redundancy as a measure of Q2 is recommended because it includes the key element of the path model, the structural model, to predict eliminated data points (Chin, 2010; Hair Jr & Hult, 2016, pp. 183-184)

Q2 > 0

Behavioral intention Q2 = 0.471

Use behavior Q2 = 0.151

Model fit

Standardized Root Mean Square Residual

SRMR

< .08

.059

Squared Euclidean Distance

d_ULS

< .95

.803

Geodesic Distance

d_G

< .95

.777

Incremental fit measure

NFI  

> 0.9

0.821

rms Theta

 

 

0.177

 

 

 

 

Multi-Group Analysis

Bootstrapping for MGA

The number of bootstrap samples for Multi Group Analysis should be high but must be at least equal to the number of valid observations. As a rule, 5,000 bootstrap samples is recommended (Hair Jr & Hult, 2016, p. 132)

2000 - 4000

5000

Path coefficient difference between groups

PLS-MGA (Henseler’s MGA) with 5% of significance level. 

t-value > 1.96

p-value < .05 or > .95

The path between Facilitation -> Use behavior has a marginal difference (p = .055 

 

 

 

 

Note: It is important to note that these model fit assessment criteria often not be useful for PLS SEM and must be used with caution. These criteria are in their very early stage of research and not fully understood. However, these fit statistics give researchers to estimate the quality of the model when it is a reflective model (Hair Jr & Hult, 2016). In more detail, please see this Note of Caution (https://www.smartpls.com/documentation/functionalities/model-fit)





References

Chin, W. W. (2010). How to write up and report PLS analyses. Handbook of partial least squares, 655-690.

Cohen, J. (1992). A power primer. Psychological bulletin, 112(1), 155.

Fornell, C., & Larcker, D. F. (1981). Evaluating structural equation models with unobservable variables and measurement error. Journal of Marketing Research, 39-50.

Hair, J. F., & Hult, G. T. M. (2016). A primer on partial least squares structural equation modeling (PLS-SEM): Sage Publications.

Hair, J. F., Ringle, C. M., & Sarstedt, M. (2011). PLS-SEM: Indeed a silver bullet. Journal of Marketing theory and Practice, 19(2), 139-152.

Hair, J. F., Ringle, C. M., & Sarstedt, M. (2013). Partial least squares structural equation modeling: Rigorous applications, better results and higher acceptance. Long Range Planning, 46, 1-12. doi:http://dx.doi.org/10.1016/j.lrp.2013.01.001

Henseler, J., Dijkstra, T. K., Sarstedt, M., Ringle, C. M., Diamantopoulos, A., Straub, D. W., . . . Calantone, R. J. (2014). Common beliefs and reality about PLS: Comments on Rönkkö and Evermann (2013). Organizational Research Methods, 17(2), 182-209.

Henseler, J., Hubona, G., & Ray, P. A. (2016). Using PLS path modeling in new technology research: updated guidelines. Industrial Management & Data Systems, 116(1), 2-20.

Latan, H., & Ramli, N. A. (2013). The Results of Partial Least Squares-Structural Equation Modelling Analyses (PLS-SEM). doi:10.2139/ssrn.2364191

Richter, N. F., Sinkovics, R. R., Ringle, C. M., & Schlaegel, C. (2016). A critical look at the use of SEM in international business research. International Marketing Review, 33(3), 376-404.

Vinzi, V. E., Chin, W. W., Henseler, J., & Wang, H. (2010). Handbook of partial least squares: Concepts, methods and applications: Springer Science & Business Media.

Williams, M. D., Rana, N. P., & Dwivedi, Y. K. (2015). The Unified Theory of Acceptance and Use of Technology (UTAUT): A Literature Review. Journal of Enterprise Information Management, 28(3), 443-488.

Wong, K. K.-K. (2013). Partial least squares structural equation modeling (PLS-SEM) techniques using SmartPLS. Marketing Bulletin, 24(1), 1-32.




댓글