Survey Explorer

Shows the response distribution per variable — how many people picked each level (weighted by survey weights). The simplest look at the data. Useful as a sanity check before more complex analyses, and for spotting categories that are too small to model reliably.

Each row is a one-variable logistic regression: how well does this predictor alone separate outcome=1 from outcome=0 cases? Tjur R² measures that separation. Different from a correlation matrix, which gives one statistic per pair: this fits a real model per variable, so it handles categorical levels (no equal-spacing assumption) and non-linear shapes correlation misses. It uses the same cross-validated metric (5-fold CV, threshold 0.005) that Build's Auto-Build uses to select predictors — so the ranking here previews what Auto-Build would consider.

Pairwise statistic — Pearson or Spearman — measuring linear (or rank) co-movement between every pair of selected variables. Unlike Predictor Strength, this gives you sign (positive vs negative direction) and shows predictor-predictor relationships, not just predictor-outcome. Use it to spot which predictors overlap with each other (collinearity) before building a multivariate model.

Shows how one variable's responses break down by another's — counts and percentages per cell. Different from Connections / Correlations, which compresses each pair to a single number: a crosstab keeps the full joint distribution, useful for non-monotonic patterns and presenting findings. The mode toggle below controls whether each picked variable is paired with the outcome or with every other picked variable.

Outcome × variable generates one table per selected variable (no limit). Variable × variable generates a table for every pair (C(N,2)) and is capped at 50 variables — at higher counts, request times exceed server limits.

Whose data?

By default, all respondents. Optionally restrict to a subgroup.

Which variables?