PLS-SEM for a
Customer Loyalty Survey

RESEARCH QUESTION

What the Study Set Out to Explain

This PLS-SEM example follows a marketing research team that surveyed members of a retail loyalty program. They want to know how perceived service quality, perceived value and brand image relate to the intention to buy again, and how well the model predicts that intention for members it has not seen.

The aim is prediction and explanation together, in a model with several constructs and one composite index. Because the data are a single cross-sectional survey, the paths are framed as associations between constructs, and the plan sets out in advance how each part of the model will be judged.

STUDY DESIGN AND DATA

A Survey With Formative and Reflective Constructs

A cross-sectional online survey with around 350 usable responses. Service quality, perceived value and repeat-purchase intention are measured by multi-item reflective scales on seven-point agreement items. Brand image is a formative index built from three components (staff, convenience and physical environment), each contributing a distinct part of the concept. The sample was checked against the largest number of paths pointing at one construct, using a power-based minimum.

METHOD AND ASSUMPTIONS

Why Partial Least Squares, and What Must Hold

Partial least squares suits this study because the aim is prediction with a model that includes a formative index, which covariance-based SEM handles less readily. The measurement model is evaluated first: loadings, composite reliability and average variance extracted for the reflective constructs, and the heterotrait-monotrait ratio (HTMT) against a prespecified threshold for discriminant validity. For the formative index, collinearity is checked with variance inflation factors and the outer weights are examined with their bootstrap intervals, with a redundancy analysis against a global single item for convergent validity.

The plan fixes a large number of bootstrap resamples with a recorded seed, and PLSpredict against a linear benchmark to judge out-of-sample predictive power. The physical environment component has a small outer weight whose interval includes zero, but its loading is substantial and the index is defined to include it, so it is kept as planned and both values are reported. Predictive power is judged per indicator, and the sample is not treated as a random draw from all program members.

OUTPUT AND INTERPRETATION

How the Results Are Read

The output reports loadings, reliability, AVE and HTMT with bootstrap intervals, then variance inflation factors and outer weights for the formative index. The structural model shows path coefficients with 95% bootstrap confidence intervals and the variance explained in the outcome. Each path is described as a cross-sectional association between constructs, with the interval showing how precisely it is estimated.

PLSpredict compares the model's indicator-level prediction errors with those of the linear benchmark, so the reader can see whether the model predicts better than a simpler alternative, not only whether paths are significant. Sensitivity checks examine the effect of the retained physical environment component and of an alternative treatment of the formative index. The write-up states what the model supports and where it cannot speak to cause.

REPORTING

What the Researcher Received

A methods and results section in journal format, with tables in APA 7 style and a short justification of each modeling decision. The constructs, the interpretation and the conclusions remain with the research team.