How to Add Control Variables in SmartPLS3? (See Description)

Adding control variables in SmartPLS3 can change whether demographic predictors look statistically significant—and that shift matters because it reveals potential confounding. In this walkthrough, gender and age are introduced as control variables to test whether they distort the relationships among the model’s endogenous constructs, specifically OC and CC (collaborative culture). Gender is treated as a categorical variable with two categories (male, female), so it’s entered directly as a latent variable without creating dummy variables. Age is treated as a continuous variable, also added as a latent variable, then the model is bootstrapped to assess significance.

With gender and age included, the results show no effect from gender, while age produces a significant effect. To understand whether age is acting as a confounder, the analysis is repeated without any control variables. The comparison is done using R-square values and the significance of the paths. When controls are removed, the R-square for OC drops slightly (from 0.495 with controls to 0.482 without), indicating that age contributes explanatory power. More importantly, age’s significance persists: age remains influential even when it is not included as a control variable. That persistence suggests age is not merely a nuisance variable—it genuinely affects the endogenous outcome, so it carries confounding influence.

The walkthrough then compares how the model’s key statistics behave when controls are added versus removed. Even when the beta coefficients and overall relationships remain broadly similar, the p-values can change. The takeaway is nuanced: significance levels may shift with or without controls, but the substantive effect of age on the endogenous construct remains. In other words, adding age as a control variable alters statistical signaling more than it alters the underlying direction of relationships.

Finally, the tutorial addresses a common complication: control variables with more than two categories. Using job rank as the example, it explains that SmartPLS requires dummy-variable coding for multi-category categorical predictors. For job rank with three levels—Junior, Middle, and Senior—three dummy indicators are conceptually created, but only two are actually added to the model so one category serves as the reference group (Junior). Middle is coded as 1 for Middle employees and 0 otherwise; Senior is coded as 1 for Senior employees and 0 otherwise. After bootstrapping with these dummy controls included, the results show no significant impact from Middle or Senior job rank on the endogenous variables. That means job rank does not produce a confounding effect, so there’s no strong statistical reason to keep it as a control.

Overall, the process is practical: add candidate controls in SmartPLS3, bootstrap to test significance, rerun the model without controls, and compare R-square and path significance to judge whether the control variable changes the relationships in a meaningful way. The tutorial’s examples show both a case where age behaves like a confounder and a case where job rank does not.