Every method AddMaple runs.

Pearson chi-square on a contingency table, with Cramér’s V as the effect size. With a weight column, the test uses weighted cell totals and the p-value is approximate for frequency weights.

Each category is collapsed against all remaining categories and tested on its own. This shows which levels carry a categorical association, beyond the omnibus chi-square.

A cell proportion is compared with the total, with its complement, or with another column or row. Pairwise column and row contrasts use a two-proportion z-test. Cohen’s h is the effect size for a difference in proportions.

The standardized residual shows how far a cell sits from the count expected under independence. Shading in a banner uses these residuals together with the adjusted p-values.

Categories that differ after Holm adjustment are marked with letters, so a stakeholder table can show which columns differ without a matrix of p-values.

Cell tests and pairwise contrasts in a banner or crosstab are adjusted with Holm’s step-down procedure. That is the multiple-comparison correction used on these tables.

Compares two group means without assuming equal variance. The effect size is Cohen’s d, using the pooled sample standard deviation.

An F test for a difference among group means. Weighted runs use Kish effective sample size for the within-group variance and the error degrees of freedom, and are marked approximate. The effect size is eta-squared.

A rank-based omnibus test. On unweighted data it is used when at least two groups have fewer than four observations and the raw samples are available. The reported effect size is an eta-squared derived from the H statistic.

Pairwise comparisons of group means after a one-way ANOVA, using the studentized range. These pairs are produced for unweighted ANOVA.

Pairwise letters for group means inside a numeric pivot, using Fisher’s LSD with Holm adjustment.

Results carry an effect size with the test: Cramér’s V, Cohen’s d, Cohen’s h, eta-squared, or a correlation. Those magnitudes are also banded as weak, moderate, or strong so a long list of tests can be ranked and colored.

Tables report bases, unweighted and weighted counts, percentages, and an index versus the total. Numeric columns report means and standard deviations, plus medians, quantiles, and box plots where the chart calls for a distribution.

The column you choose is the banner: Ranked differences by that column, as charts or as a table, sorted by significance or left in data order. Each other question is tested with the method for those two column types and ordered by effect size. A coded open-end theme is a categorical column and is ranked with the rest.

For each answer, a two-proportion z-test compares that segment’s share with the share among everyone else. Holm adjustment is within that answer, across the segments. A shaded cell is higher or lower than the rest of the sample. Cohen’s h is the effect size.

The count expected if the banner and the question were independent, from the chi-square on that table. The standardized residual, with a Holm-adjusted p-value within each answer, shows whether that segment’s answer is higher or lower than expected.

Pick the baseline in the control, or click a column header. Total can be that baseline. Each other segment is compared with it on each answer, using a two-proportion z-test. Holm adjustment is within that answer and that baseline. Cohen’s h is the effect size.

Each pair is tested with the method in the table above and ranked by effect size and significance. The list is a shortlist of associations for a later driver model or a focused table.

Used when both numeric axes pass a normality heuristic: absolute skew under 1 and absolute excess kurtosis under 2. Weighted Pearson correlations are approximate.

Rank correlation for numeric pairs that fail the skew and kurtosis check, including skewed scores and ranks.

A linear model with both L1 and L2 penalties, so many mixed-type predictors can be ranked without keeping every one. Coefficients keep a sign and a magnitude. The penalty is chosen along a lambda path.

The same regularized family for an ordered outcome such as an agreement scale or NPS, where the category order matters. It needs at least three ordered categories.

A regularized model for an unordered outcome with at least three categories. Coefficients are class-specific, so a Likert or multiclass target can show a different driver pattern for each class.

Fitted by iteratively reweighted least squares. Coefficients are signed log-odds, with an odds-ratio reading and a pseudo-R². Key drivers use it for a two-class outcome. The regression tool uses it for a numeric predictor and a binary outcome, with a fitted curve.

A classification forest of CART trees split on Gini impurity. Importance is the mean decrease in Gini across splits, and fit is reported with out-of-bag accuracy. Split importance rises for predictors with many distinct values. Direction comes from the coefficient models.

Ordinary least squares in the regression tool. The result shows the fitted line, R², the coefficient, and a confidence interval.

Density-based clustering for mixed or categorical data. The number of groups comes from the data. Rows outside a dense region are labelled noise.

A finite mixture fit by expectation-maximization. Items can be categorical, ordinal treated as categorical, boolean, or multi-select indicators, under a local-independence assumption. Class counts are compared with AIC, BIC, and entropy. Weighted BIC uses Kish effective sample size.

HDBSCAN finds dense regions, then k-medoids fits the number of groups you set. The density labels affect the starting centers. Group sizes are not forced to be equal.

Euclidean clustering with a chosen number of groups. Numeric columns can stay raw, be row-centered, or be ipsatized: row-centered and divided by the row standard deviation.

A chosen number of groups on mixed survey variables. Centers are real respondents, and the distance is Gower.

The distance behind k-medoids and the density methods. Numeric and ordinal columns are range-scaled, categorical columns contribute a match or a mismatch, and multi-select columns use Jaccard overlap.

Measures how well each respondent sits inside its group versus the neighboring group, on the distance the algorithm used. Compare scores inside one distance family: Euclidean for k-means, Gower for the others.

A permutation test of whether the fitted groups differ on the distance matrix that produced them. It uses the same features as the clustering, so it describes that fitted partition.

Agreement between a partition and refits on resamples, adjusted for chance. Reported for k-means, k-medoids, HDBSCAN, and HDBSCAN-seeded k-medoids. Latent class runs are compared with AIC, BIC, and entropy.

Each item gets a best rate and a worst rate, shrunk with a Beta prior. Utility is the difference of the two logit probabilities, centered across items. Preference share is a softmax of those utilities. Share intervals come from a Monte Carlo sample of that uncertainty. Rank by utility, or by best minus worst.

Each level records wins and exposure across the concepts where it appeared. A Beta prior shrinks the win rate. Part-worths are those logits, centered within the attribute. Importance is the range of part-worths, also shown as a share of the total range. Scenario shares are a softmax of summed part-worths, with Monte Carlo intervals.

Searches item combinations from a multi-select or a set of binary reach columns. A respondent counts toward a combination when they selected at least one of its items. Weights apply to that reach. You choose the combination sizes; the search stops at an evaluation cap.

Correlation PCA. Columns are standardized, the correlation matrix is decomposed, and components are kept with their loadings and the variance they explain. Optional Varimax rotation keeps the components orthogonal and simplifies which questions load where.

0–6 are detractors, 7–8 are passives, and 9–10 are promoters. NPS is the percent promoters minus the percent detractors.

Calculated columns for CSAT on a 5- or 7-point scale, CES on a 5- or 7-point scale, UMUX-Lite from two items, and UMUX from four items.

Scale polarity produces top-box, bottom-box, and net scores, including top-2 and bottom-2 boxes. A 0–10 scale is recognized as NPS. Just-about-right batteries can be charted as JAR scales.

Coefficient alpha for the items in a scale, with a reliability band, before that score is used as a driver, a segment input, or a headline metric.

Numeric calculated columns for indices and other composites. The formula stays on the project with the columns it reads.

Iterative proportional fitting adjusts respondent weights toward demographic target proportions. The result reports iterations, remaining margin error, the weight range, Kish effective sample size, and weighting efficiency.

Effective n is the square of the sum of weights, divided by the sum of squared weights. Weighted chi-square uses weighted cell totals and is approximate. Weighted ANOVA uses this effective n for variance and degrees of freedom and is approximate.

Mental penetration is the share of respondents who associate a brand with at least one category entry point. Mental market share, network size, mental advantage, and share of mind summarize how broad and how strong those associations are. The CEP-by-brand matrix compares brands across entry points.