Categorical Predictor Variables Using SmartPLS/Dummy Variables Regression Model in SMART-PLS3
Based on Research With Fawad's video on YouTube. If you like this content, support the original creators by watching, liking and subscribing to their content.
Convert nominal categorical predictors (e.g., Country) into dummy variables before using them in SmartPLS regression models.
Briefing
Categorical predictor variables in SmartPLS must be converted into dummy variables before they can be used in regression-style models—otherwise SmartPLS/PLS bootstrapping can fail due to singular matrix problems. The core workflow starts with a nominal variable like Country (e.g., China, Pakistan, Italy), which has no natural order and therefore can’t be treated like metric inputs such as CSR, loyalty, or satisfaction. In practice, a three-category nominal predictor becomes two dummy variables, because the third category acts as a reference group for comparison.
The transcript walks through creating those dummies in SPSS: using Transform → Create Dummy Variables, selecting Country, and assigning a root name (e.g., “country”) so SPSS generates one column per category (Country 1, Country 2, Country 3). In the dataset, each respondent is coded with 1 for their category and 0 for the others—for example, a respondent from Italy has 0s in the China and Pakistan dummy columns. The resulting file is saved as CSV and imported into SmartPLS.
In SmartPLS, the model is then set up to test whether Country affects perceptions toward CSR. The transcript highlights a common mistake: linking all dummy columns (China, Pakistan, Italy) to CSR without designating a reference category leads to a singular matrix error. The reason is technical and straightforward: one or more dummy columns end up with zero variance or perfect collinearity because the model has no baseline category to compare against. The fix is to remove one category so the remaining dummies compare their effects against a single reference group. When Italy is removed and used as the reference, bootstrapping runs successfully.
With Italy as the baseline, the results show that China has a positive coefficient for CSR perceptions relative to Italy, indicating improved perceptions. Pakistan shows a negative coefficient relative to Italy, indicating lower perceptions. The transcript further notes that these differences are significant based on the bootstrapping output.
A second example uses another categorical predictor: job rank (Junior, Middle, Senior). Here, Junior is chosen as the reference category. The model tests effects on three outcomes—collaborative culture, reliability, and organizational commitment. Middle-level employees show little to no impact on collaborative culture (low coefficient). Senior-level employees, by contrast, have a significant positive impact on collaborative culture compared with juniors. Senior-level employees also significantly increase organizational commitment relative to the reference group, while showing no meaningful influence on reliability.
Overall, the key takeaway is that nominal categorical predictors require dummy coding plus an explicit reference category; doing both prevents singular matrix failures and makes the coefficients interpretable as comparisons against the baseline group.
Cornell Notes
Nominal categorical predictors (like Country with China, Pakistan, Italy) can’t be used directly in SmartPLS regression models because they lack order and behave like non-metric inputs. The solution is dummy coding: a three-category variable becomes two dummy variables, with the third category serving as a reference group. If all dummy columns are included without a reference, SmartPLS bootstrapping can fail with a singular matrix error due to zero variance/perfect collinearity. Once one category is removed as the baseline, bootstrapped coefficients become interpretable as differences in the outcome (e.g., CSR perceptions) relative to that reference. The same approach applies to other categorical predictors like job rank (Junior as reference) when testing effects on collaborative culture, reliability, and organizational commitment.
Why can’t a nominal variable like Country be used directly as a predictor in SmartPLS regression models?
How many dummy variables are needed for a nominal predictor with three categories, and why?
What causes the singular matrix problem when all dummy categories are linked to the outcome?
In the Country → CSR example, how should the sign of the coefficient be interpreted?
How does the job rank example use a reference category, and what were the main outcome effects?
Review Questions
- If a nominal predictor has k categories, what is the general rule for how many dummy variables should be included in SmartPLS, and what role does the omitted category play?
- Describe the specific modeling error that occurs when no reference category is used for dummy variables, and explain the underlying statistical reason.
- In a dummy-coded model with a baseline group, how do you interpret a positive vs. negative coefficient for a non-baseline category?
Key Points
- 1
Convert nominal categorical predictors (e.g., Country) into dummy variables before using them in SmartPLS regression models.
- 2
For a k-category nominal variable, include only k−1 dummy variables and omit one category as the reference group.
- 3
Including all dummy categories without a reference can trigger singular matrix errors due to perfect collinearity/zero variance.
- 4
Interpret dummy coefficients as differences in the outcome relative to the omitted reference category (sign indicates direction).
- 5
Use bootstrapping in SmartPLS to assess whether category differences are statistically significant.
- 6
Apply the same dummy + reference approach to other categorical predictors like job rank when testing effects on multiple outcomes.