Statistical Associates Publishers

## Partial Least Squares Modeling: 10 Worst Pitfalls and Mistakes

1. Not testing the measurement model.
There is no point in testing the structural (inner) model until the measurement (outer) model is upheld.

2. Testing only one model
For any given set of data it is possible more than one model will have "good fit". It is better to compare models to determine which has better fit. Like almost all multivariate procedures, PLS results may differ markedly if previously-omitted important causal variables are added to the model, though PLS is less sensitive to inclusion of spurious causal variables correlated with indicator variables.

3. Choosing a formative model when the indicators are reflective, or vice versa.
Computed results are erroneous is the indicators do not conform to the model. In reflective models, indicators are reflections of their component and as they all reflect the same thing, they should intercorrelate highly. In formative models, indicators additively compose the component and may well not intercorrelate highly.

4. Using the wrong fit (quality) measures for the type of model.
Model fit measures differ between reflective and formative models as do interpretations of the same measure. Cronbach's alpha, for instance, conforms to the assumptions of reflective models but not formative models.

5. Not testing the data for unobserved heterogeneity .
A global PLS solution is misleading if the data are segmented. The researcher must select the proper method for testing and dealing with this common complexity.

6. Treating bootstrapped significance as the same as conventional parametric significance testing based on random sampling..
PLS coefficients lack a known distribution, making the usual parametric significance testing impossible. Therefore bootstrapped significance is used as an alternative. However, the meaning of significance coefficients is not the same for the two methods. Bootstrapped significance assures generalization to the dataset at hand, not necessarily to the target population from which the data were sampled.

7. Assuming sample size does not matter. .
PLS, unlike SEM, can be computed for small samples, even samples smaller than the number of variables. However, it remains true that small samples may be unrepresentative and large samples increase the stability and reliability of model coefficients.

8. Interpreting PLS models and coefficients the same as SEM models and coefficients .
PLS is a variance-based approach using principle components. SEM is a covariance-based approach using common factors. Coefficients and their interpretation will differ for the same causal model.

9. Not examining residuals. .
In a well-fitting model, tesiduals should be uncorrelated with independent variables and should be random normal, as in other predictive procedures

10. Not meeting the assumptions of partial least squares modeling. .
Our book, listed below, enumerates many assumptions of PLS, clearly listed in the "Assumptions" section.