Regression, PCA & multivariate modelling

What is the estimated relationship — and what structure sits behind the batteries?

When the question moves beyond crosstabs and one-at-a-time rankings, AddMaple keeps explicit multivariate modelling inside the live project — with fitted charts, inspectable coefficients, and save-to-Insights workflows. You do not need a separate stats stack to move from tables to models.

Linear and logistic regression

Build a focused regression model the researcher chooses and reviews directly.

• Linear regression — numeric predictor and numeric outcome; scatter plot with regression line, R², coefficient, confidence interval, and plain-language interpretation.
• Logistic regression — numeric predictor and binary outcome (two-category tag, boolean, or assigned-scale column); fitted curve, log-odds, pseudo R², and odds-ratio interpretation.

Outlier toggles, jitter, and chart views make the fit reviewable before it becomes a recommendation. Save results to Insights or dashboards so stakeholders can revisit the estimate without rerunning it.

Multivariate drivers with Elastic Net and Random Forest

When many candidate factors could explain an outcome, use key driver analysis in the same project:

• Elastic Net — numeric outcomes with regularised coefficients across many predictors; direction and magnitude stay interpretable.
• Logistic Elastic Net — binary outcomes with multivariate regularisation.
• Ordinal Elastic Net — ordered scales such as NPS or agreement grids where category order matters.
• Multinomial Elastic Net — unordered multiclass outcomes with class-specific driver evidence.
• Random Forest — complex mixed-type patterns with out-of-bag fit when relationships are non-linear.

Automatic routing picks a target-appropriate model; analysts can override when the study demands a specific approach. Related-column discovery surfaces plausible drivers before you run the multivariate model.

PCA and dimension reduction

Run principal component analysis on numeric and opinion-scale batteries to explore latent structure and reduce dimensionality beyond one-at-a-time crosstabs.

• Select correlated attitude, rating, or sensory columns and review loadings and variance explained.
• Optional Varimax rotation for simpler, more interpretable dimensions.
• Use component maps and loadings to decide how to tell the story — or feed components into segmentation when the study calls for it.

PCA is exploratory: it summarises co-movement in the data, not causal drivers. For outcome ranking across many candidates, start with key driver analysis.

Choose the method that answers the question.

• Which of many factors matter most for an outcome? Start with key-driver analysis — Elastic Net, ordinal or multinomial Elastic Net, or Random Forest depending on the target.
• What is the estimated relationship for a focused predictor and outcome? Use linear or logistic regression with a reviewable scatter or curve.
• What underlying dimensions sit behind this battery? Use PCA.
• Which items or concepts win preference tasks? Use MaxDiff, TURF & Conjoint.

“What happens to spend as tenure rises?” is a focused regression question. “Which of twelve experience drivers best explain NPS?” is a multivariate key-driver question. “What are the main dimensions in this attitude grid?” is a PCA question. Keeping those workflows distinct preserves statistical clarity.

Keep the estimate attached to its assumptions.

Regression and driver models make relationships easier to inspect, but they do not make an observational result causal by themselves. Review predictor and outcome definitions, model fit, outliers, bases, and study design before carrying a coefficient into a recommendation. Saving the model to Insights or a dashboard keeps that context available when the result is revisited.