PLS-SEM and reflective constructs: A response to recent criticism and a constructive path forward

The authors begin by acknowledging their previously published article in Industrial Marketing Management (IMM), focusing on PLS-SEM guidance for business marketing researchers. The article's high citation rate highlights the importance of addressing common errors in PLS-SEM applications and exploring advanced techniques. They then address a commentary by Henseler et al. (2025) which criticized their work, raising two main arguments: 1) PLS-SEM is unsuitable for reflectively measured constructs due to biased parameter estimates, and 2) the assessment criteria for evaluating reflective measurement models are also biased. This response paper aims to demonstrate why these concerns are unfounded.

The authors extensively review the literature to distinguish between model design and model estimation. They draw on Rigdon's (2012) proxy framework, differentiating between conceptual variables (proxies representing constructs in statistical models) and indicators linked to proxies through mathematical operations. This framework highlights the validity gap between the proxy and the concept, a notion also discussed by Rossiter (2002, 2011a, 2011b). The concept of metrological uncertainty (Rigdon et al., 2020; Rigdon & Sarstedt, 2022) is introduced, quantifying the range of a measured quantity's true value. This uncertainty impacts all elements of the research process, emphasizing that measurement approximates the underlying concept. Sarstedt et al. (2016) are cited to highlight the distinction between a measurement model's conceptualization/operationalization and the data-generation process, emphasizing that reflective or formative operationalization is a conceptual decision grounded in measurement theory. The authors discuss how highly correlated indicators might support reflective specification even if the underlying data follows composite model logic, referencing the flexible nature of attitudes (Zaller & Feldman, 1992; Fazio et al., 1984; Regan & Fazio, 1977; Stern et al., 1995). The use of formative measurement models in CB-SEM further illustrates the separation of theoretical model specification and model estimation. The authors contrast their perspective with Henseler et al.'s (2025) view, which they see as a purely statistical practice lacking measurement-theoretic considerations. They cite Rigdon, Sarstedt, and Ringle (2017) to explain the confusion stemming from researchers' backgrounds and philosophical positions in science. Finally, the authors discuss the consequences of incorrect assumptions about the data-generating process, citing Sarstedt et al. (2016) and Cho, Sarstedt, and Hwang (2022) to show PLS-SEM's robustness compared to CB-SEM under model misspecification. Deng and Yuan (2023) are also cited to support the advantages of path analysis with composite scores for prediction and classification.

The authors first address Henseler et al.'s (2025) claim that reflective measures equal common factor models, questioning the validity of this assumption. They then explore the consequences of this assumption for measurement model results and validation. The assessment of reflectively measured constructs in PLS-SEM is outlined, including indicator reliability, internal consistency reliability (Cronbach's α, composite reliability ρc, reliability coefficient ρA), convergent validity (AVE), and discriminant validity (HTMT). The authors analyze which criteria would be affected by potentially inflated loadings from PLS-SEM. They argue that HTMT and ρA are unaffected, while Cronbach's α is too conservative and ρc too liberal. The impact on indicator reliability and AVE is acknowledged, as PLS-SEM might return higher values than CB-SEM due to not dividing variance into common and unique variance (Guenther et al., 2023; Sarstedt et al., 2016). The authors discuss studies (Cho, Sarstedt, and Hwang, 2022; Schuberth, Hubona, et al., 2023; Dash and Paul, 2021) that evaluated the extent of loading inflation in PLS-SEM when estimating common factor models. These studies indicate that the inflation is relatively small and decreases with larger sample sizes and more indicators, due to PLS-SEM's consistency at large characteristic (Hui & Wold, 1982; Schneeweiß, 1993). The authors analyze Sarstedt et al.'s (2022) review of PLS-SEM use in marketing journals to evaluate the practical impact of potentially inflated AVE values. They find that even with a loading inflation correction, most constructs still meet the convergent validity threshold. They cite Hair et al. (2022) to support the idea that differences between PLS-SEM and CB-SEM estimates are small when measurement models meet minimum standards (four or more indicators, loadings ≥0.70). The authors counter Rönkkö et al.'s (2023) critique of AVE, referencing Hair et al. (2024a) who showed that this critique is based on selective reporting and outdated guidelines. The authors emphasize the distinction between a theoretically conceptualized reflective measurement model and its statistical estimation using empirical data. They contrast this with Henseler et al.'s (2025) view and highlight inconsistencies with the existing literature. The authors use the example of moderation in regression analysis to show the distinction between conceptual and statistical models, and they extend this analogy to PLS-SEM. They discuss various methods for estimating reflectively operationalized conceptual variables, including CB-SEM, PLS-SEM, PLSC-SEM/PLSe-SEM, GSCA, IGSCA, factor score regression, and sum score regression, noting the different statistical models and assumptions involved. They also discuss the inherent uncertainty about the data-generating process and the resulting validity gap between proxies and conceptual variables. They argue that PLS-SEM is a safer choice when the underlying model type is unknown, supported by Rigdon (2024) who shows consistency between RCA, PLS path modeling, and GSCA for factor model data. Finally, they propose a multimethod approach to SEM, using different methods as tools in a toolbox. They discuss the use of robustness tests and the importance of focusing on the robustness of inferences rather than differences in estimation results across methods. They address potential sources of differences in results across methods: different statistical models, data imperfections, and analytical decisions. They also discuss the complementary nature of CB-SEM (model fit) and PLS-SEM (predictive power).

The key findings revolve around the refutation of Henseler et al.'s (2025) criticisms of PLS-SEM for reflective measurement. The authors demonstrate that the assumption of equivalence between reflective measurement and common factor models is incorrect. They show that even if this assumption were true, the consequences for measurement model results and validation are minor based on various studies showing minimal bias and inflation in PLS-SEM estimates under common factor model conditions. The analysis of AVE values from a large body of published research (Sarstedt et al., 2022) demonstrates that the vast majority of constructs maintain convergent validity even when accounting for potential loading inflation. The authors conclude that PLS-SEM is a robust and reliable method for handling reflective measurement models, particularly given the uncertainty inherent in real-world data about the data-generating process. They emphasize the importance of considering different SEM methods as tools within a methodological toolbox rather than mutually exclusive options, advocating for a multimethod approach to enhance robustness and address methodological uncertainty. The authors further find that differences between results obtained using various methods (CB-SEM with its various estimators, PLS-SEM, PLSC-SEM/PLSe-SEM, GSCA, and IGSCA) are less pronounced for reflectively measured constructs compared to formatively measured constructs from a composite population. The robustness of findings across different methods is highlighted as a crucial aspect of a multimethod approach, prompting researchers to investigate discrepancies in results and refine models, data, or methods as needed. A comparison of CB-SEM's model fit assessment and PLS-SEM's predictive power assessment highlights the utility of combining both approaches. The study suggests that PLS-SEM’s robustness and less restrictive requirements could even be used to identify potential issues in CB-SEM model specification and estimation.

The paper's findings directly address the research question by refuting the criticisms of PLS-SEM for reflective constructs. The authors successfully demonstrate the limitations of the critics' assumptions and highlight the robustness of PLS-SEM under various conditions. The significance lies in providing a strong defense of PLS-SEM's utility, while simultaneously advocating for a more comprehensive and pluralistic approach to SEM methodology. The authors' call for a multimethod approach is a significant contribution, acknowledging the inherent uncertainties in real-world data and emphasizing the importance of cross-validation and robustness testing. The findings are relevant to a broad range of disciplines using SEM, promoting more rigorous and nuanced interpretations of results. The discussion of the validity gap and metrological uncertainty provides important context for understanding the inherent limitations of any measurement method. The practical implications of the findings are substantial, providing researchers with a more robust framework for conducting and interpreting SEM analyses.

The authors conclude that the criticism of PLS-SEM for reflective measurement models is unfounded. They reject the "either-or" dichotomy and promote a "both-and" perspective, advocating for a multimethod approach that combines CB-SEM and PLS-SEM to leverage their respective strengths in model fit assessment and predictive power. Future research should focus on refining the understanding of conceptual and statistical models, improving construct definition and estimation, exploring flexible cutoff values for evaluation criteria, and addressing the validity gap between conceptual constructs and their statistical proxies. The adoption of open science initiatives is also recommended to improve transparency and replicability.

The paper primarily relies on a review of existing literature and simulations to support its arguments. While the authors cite multiple studies, it might be beneficial to conduct original empirical research to further strengthen the claims regarding the relative merits of PLS-SEM and CB-SEM under specific conditions. Furthermore, the discussion of a multimethod approach is conceptual, and providing detailed practical guidelines for implementing such an approach could further enhance the paper's contribution. The paper's focus on business marketing research might limit the generalizability of some findings to other fields, though the authors do mention applications across multiple disciplines.