This article by Ned Kock, published in October 2015, offers a comprehensive discussion and practical solution for common method bias (CMB) within the domain of structural equation modeling (SEM) specifically employing the partial least squares (PLS) method (PLS-SEM). SEM, rooted in path analysis developed by Wright, relies on models that depict causal relationships among variables, particularly latent variables measured indirectly through indicators, often obtained via questionnaires with Likert-type scales.
The paper defines CMB as a phenomenon originating from the measurement method itself, rather than the true causal network among latent variables in an SEM study. For instance, general instructions on a questionnaire or the implicit social desirability linked to certain answers can cause indicators to share a common variation. Mathematically, the article illustrates how this common method variation, represented by a standardized variable M and its weight ωm, introduces a shared component (ωm*M) into the derivation of each indicator from its latent variable. This shared variation has the critical effect of artificially increasing the level of collinearity among latent variables, which, in turn, predictably inflates path coefficients. This inflation of path coefficients is a significant concern for researchers as it can lead to Type I errors (false positives), suggesting a relationship exists when it does not.
To thoroughly illustrate and test his claims, Kock employed a Monte Carlo simulation to create two datasets of 300 rows each, simulating questionnaire responses: one explicitly contaminated with CMB and one without. The simulation used an illustrative model inspired by e-collaboration research, featuring three latent variables—collaborative culture (F1), e-collaboration technology use (F2), and competitive advantage (F3)—each measured by six indicators. In the simulated “true” model, all path coefficients were set at .45 and all indicator loadings at .7. For the contaminated dataset, the common method weight was set at .6. The analysis was performed using WarpPLS software, version 5.0, utilizing the PLS Mode A algorithm with a path weighting scheme.
A key finding of the article is the ineffectiveness of traditional confirmatory factor analysis (CFA) in identifying CMB.
- Convergent Validity: This is assessed by how strongly indicators load on their corresponding latent variables, with loadings typically expected to be above .5. The study found that both the CMB-contaminated and uncontaminated models passed this test, as CMB artificially inflates loadings, paradoxically leading to an artificial increase in perceived convergent validity.
- Discriminant Validity: This is typically met when the square root of a latent variable’s average variance extracted (AVE) is greater than its correlations with other latent variables. Again, both models, with and without CMB, displayed acceptable discriminant validity. The article explains that while CMB increases correlations among latent variables, it simultaneously increases their AVEs, thereby undermining the ability of this check to identify the bias. In summary, the article demonstrates that neither a convergent nor a discriminant validity test, commonly used in CFA, effectively identifies common method bias.
As an effective alternative, the paper proposes and demonstrates a practical approach for identifying CMB based on Variance Inflation Factors (VIFs) generated through a full collinearity test. This full collinearity test, proposed by Kock & Lynn (2012), provides a comprehensive assessment of both vertical (predictor-predictor) and lateral (predictor-criterion) collinearity within a model. The procedure, automated by WarpPLS, generates VIFs for all latent variables. A VIF greater than 3.3 is suggested as an indication of pathological collinearity and, crucially, as an indication that a model may be contaminated by common method bias. In the simulation results, the CMB-contaminated model clearly showed one latent variable (F3) with a VIF of 3.720, exceeding the 3.3 threshold, while all VIFs in the uncontaminated model remained below this level (e.g., F3 VIF was 1.739). This success is consistent with the mathematical understanding that CMB introduces common variation, thereby increasing the overall level of collinearity in the model, which is then reflected in higher full collinearity VIFs.
The article also touches upon the ongoing debate among methodological researchers regarding CMB and suggests that if the problem is real, prevention strategies (e.g., from Podsakoff et al., 2003) and post-hoc remedies for collinearity (e.g., indicator removal, re-assignment, latent variable removal or aggregation, hierarchical analysis) discussed by Kock & Lynn (2012) can be employed if CMB is detected. It notes that VIFs tend to increase with model complexity and discusses how different PLS-SEM algorithms might affect VIFs. While classic PLS-SEM algorithms can attenuate path coefficients and reduce collinearity “too much,” newer factor-based PLS-SEM algorithms (which incorporate measurement error) are expected to yield slightly higher, more accurate VIFs. Consequently, when using algorithms that incorporate measurement error, a higher VIF threshold, such as 5, might be more appropriate.
Ultimately, the article’s goal is to equip empirical researchers with a practical and straightforward methodological solution to assess the overall quality of their measurement frameworks. While the discussion is illustrated with e-collaboration research outputs from WarpPLS, the presented approach is broadly applicable to any field utilizing path analysis and SEM.
Reference: Kock, N. (2015). Common method bias in PLS-SEM: A full collinearity assessment approach. International Journal of e-Collaboration, 11(4), 1-10.
Note: If you find evidence of common method bias (CMB), there are various methods to address it, either to prevent its occurrence during data collection or to deal with it post-hoc in the analysis.
Prevention Strategies:
• A seminal source by Podsakoff et al. (2003) provides a number of suggestions on how to avoid the introduction of common method bias during data collection. These strategies aim to prevent the bias from affecting the data in the first place.
Post-Hoc Remedies (if CMB is detected): If CMB is identified through methods like the full collinearity test, steps can be taken to eliminate or reduce it. The article specifically refers to the steps discussed by Kock & Lynn (2012) for dealing with collinearity, which are considered an obvious choice given the focus on collinearity in CMB detection. These steps include:
• Indicator removal: Removing specific indicators from the model.
• Indicator re-assignment: Re-assigning indicators to different latent variables.
• Latent variable removal: Removing an entire latent variable from the model.
• Latent variable aggregation: Combining multiple latent variables into a single one.
• Hierarchical analysis: Employing a hierarchical model structure.