Sampling weight adjustments in partial least squares structural equation modeling: guidelines and illustrations

Abstract

Applications of partial least squares structural equation modelling (PLS-SEM) often draw on survey data. While researchers go to great lengths to document reliability and validity statistics that support the generalisability of their findings, they often overlook or ignore a more fundamental issue related to data analysis—the representativeness of their sample. Addressing this concern, the present paper offers guidelines for using the weighted PLS-SEM (WPLS-SEM) algorithm to apply sampling weights in the model estimation. The results of the WPLS algorithm and the traditional PLS algorithm are then compared using a marketing research model. The findings show that researchers should routinely consider the procedure of the WPLS algorithm when using the PLS technique for assessment. The WPLS algorithm is a useful and practical approach for achieving better average population estimates in situations where researchers have a set of appropriate weights. This paper substantiates the use of the WPLS algorithm and provides business researchers and practitioners with the proper guidelines to assess, report, and interpret PLS-SEM results. It also illustrates that the use of the WPLS algorithm produces different inference test results in the structural model and different predictive relevance results. Thus, the study contributes to the advancement of PLS-SEM applications.

Keywords:

• marketing
• sampling weights
• post-stratification weights
• weighted PLS (WPLS)
• PLS-SEM

Acknowledgements

The authors thank Marko Sarstedt for his guiding comments and ideas regarding an earlier version of this manuscript.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 The remainder of this article will only use the term ‘sampling weights’.

2 The composite scores have a trait of zero mean and standard deviation of one.

3 When using Mode A (i.e. correlation weights), the bivariate correlation between each indicator and the construct determines the outer weights. In contrast, Mode B (i.e. regression weights) computes indicator weights by regressing each construct on its associated indicators (Hair et al., 2017b).

4 Researchers can draw on the IF function in Microsoft Excel or use a statistical software program such as SPSS (Sarstedt & Mooi, 2019).

5 Cheah et al. (2018) use the same model and data in their comparative study on the use of single versus multiple items in PLS-SEM-based redundancy analyses.

6 The weighting can also be initiated in the context of the consistent PLS-SEM algorithm (Dijkstra & Henseler, 2015).

7 Note that Geweke and Meese’s (1981) criterion, which Sharma et al. (2019a, 2019b) also identified as suitable, is not applicable because our model is saturated.

8 The PLSpredict result for the PLS-SEM algorithm was estimated using SmartPLS (Ringle et al., 2015) while the PLSpredict result for the WPLS-SEM algorithm was estimated using the package https://github.com/ISS-Analytics/pls-predict for the statistical software R (Shmueli et al., 2016).