16. SEMinR Lecture Series | Evaluating Formative Measurement Model | Introduction

Formative measurement models in SEMinR (PLS-SEM) require a different evaluation workflow than reflective ones, and the practical difference starts at model specification. When constructs are formative, they must be declared as such—typically by setting the measurement mode to “mode B” in the composite specification—so the PLS path model estimates formative indicator weights rather than treating indicators as interchangeable reflections of an underlying latent variable. In contrast, reflective constructs use “mode A” by default and are evaluated with reliability and validity checks.

The lecture frames the overall goal: evaluating formative measurement models using criteria such as convergent validity (for the relevant parts of the model), indicator collinearity, statistical significance, and the relevance of indicator weights. It then walks through an applied example with three formative constructs—Vision development, and Reward—measured by multiple items, which jointly predict a reflective construct called Collaborative culture. This mixed measurement setup matters because only the reflective construct receives the standard reflective measurement assessment (reliability and validity), while formative constructs are assessed through weight-related diagnostics.

On the implementation side, the workflow begins with loading the dataset in R and defining the measurement model. For formative constructs, the composite function is updated with a “weights” argument that specifies “mode B” (or explicitly sets “mode A” for reflective constructs and “mode B” for formative ones). The example emphasizes that omitting this step leaves constructs treated as reflective by default, which would be incorrect for formative constructs.

Next comes model estimation. The PLS path model is estimated by calling the appropriate SEMinR estimation function (using the specified measurement model and structural model). The structural model is defined through the relationships between constructs—here, the three formative predictors affecting the reflective outcome. After estimation, results are retrieved via a summary object (stored in a user-named variable), and the output can be explored with targeted calls (e.g., using dollar-sign accessors for specific result sections).

A key quality-control step is checking missing data and algorithm behavior. The summary output includes descriptive statistics with the number of missing observations, and the iteration diagnostics (e.g., maximum iterations set to 300). In the example, the run reports all observations valid, shows no missing values, and stops well before the maximum iteration count—signals that the estimation process is stable.

Finally, the lecture distinguishes what to evaluate for reflective versus formative constructs. The reflective construct (Collaborative culture) is assessed using reliability and validity outputs such as HTMT (heterotrait–monotrait correlations) to evaluate discriminant validity, and reliability metrics to confirm internal consistency. The formative constructs (Vision development, and Reward) are not subjected to reflective reliability/validity metrics; instead, their evaluation focuses on formative-specific criteria like indicator collinearity and the significance/relevance of their estimated weights. The session closes by previewing the next step in the formative evaluation sequence: establishing convergent validity.