Interactive Statistics
What this domain tests
The Interactive Statistics domain covers Dataiku's no-code exploration environment on a dataset: the Statistics tab, worksheets, cards, filters, sampling, and the interpretation of univariate, bivariate, correlation, PCA, clustering, and statistical-test output. The exam expects you to know how to profile a dataset quickly, isolate a segment, choose a suitable card, and avoid drawing causal or overconfident conclusions from exploratory results.
Core rules and concepts
- Worksheets and cards: A worksheet groups cards. Each card answers one question: distribution, correlation, scatter, box plot, PCA, clustering, or a statistical test. Cards can be resized and arranged for a narrative.
- Quick stats: The dataset's quick statistics show count, missing values, distinct count, min, max, mean, standard deviation, and quartiles. Use them to spot constant columns, suspicious missingness, and impossible ranges before modeling.
- Univariate analysis: Histograms and box plots reveal skew, multimodality, and outliers. A high maximum may be valid or a data-entry error; inspect records rather than deleting automatically.
- Bivariate analysis: Compare a measure across categories or against another measure. Box plots by group are useful for comparing distributions; scatter plots reveal nonlinearity and heteroscedasticity.
- Correlation: The correlation matrix summarizes linear association. It does not detect nonlinear relationships and is sensitive to outliers. Correlation is not causation.
- PCA and clustering: PCA reduces numeric dimensionality for exploration; clustering groups similar records. Both require scaling decisions and cautious interpretation of components or cluster labels.
- Statistical tests: Chi-square tests association between categorical variables; t-tests and ANOVA compare means across groups; correlation tests assess linear association. Check assumptions and sample sizes.
- Filters and sampling: Worksheet filters restrict the population; sampling controls cost on large datasets. A filtered or sampled result must not be generalized to the full dataset without checking representativeness.
Worked example
A support dataset has 12,000 tickets with resolution_hours, priority, region, first_response_minutes, and reopened. Filter to closed tickets. A box plot of resolution_hours by priority shows high-priority tickets resolving faster, but with a long tail. A t-test or ANOVA confirms the mean difference; a chi-square test checks whether reopened rates differ by priority. Correlation between first_response_minutes and resolution_hours may be positive, but it does not prove that faster response causes faster resolution. Before reporting, check missing values and whether the filter excluded a large segment.
Traps that cost marks
- Running many tests and reporting only significant results without correction.
- Treating a filtered or sampled worksheet as the full population.
- Reading correlation as causation or assuming it captures nonlinearity.
- Comparing groups with very different sample sizes or missingness.
- Using numeric codes for categories in correlation or PCA.