Data Analysis Assistant

Problem It Solves

When stakeholders ask to “look at the data,” the bottleneck is often not retrieval, but turning metric changes into defensible decisions: whether the sample size is sufficient, whether p-values are inflated by multiple comparisons, whether retention curves mix in time effects, and whether percentages can be averaged directly. Data Analysis Assistant moves these concerns into a checklist, helping analysts define the decision, the falsification condition, and data boundaries before modeling, so statistical significance is not mistaken for business significance.

How It Works

It starts by asking what decision the analysis supports: if the result points to X, what would the team do differently? Then it selects methods by question type: hypothesis tests for group differences, outputting p-value, effect size, and confidence interval; regression or correlation for prediction, with residual checks and R²; cohort analysis for behavior over time; segmentation for group profiles; and anomaly detection with contextual flags.

In the output, it prioritizes the insight over methodology, quantifies uncertainty with ranges such as 12%-18% instead of a single 15% point estimate, and states limitations plus next validation steps. It also flags common pitfalls, including Simpson’s Paradox, survivorship bias, unequal time windows, p-hacking, spurious time-series correlation, and direct averaging of percentages.

Boundaries

This skill fits scenarios where the business question, sample, and metric definitions are already clear, especially A/B testing, retention, churn, metric anomalies, and segment comparison. It is not an automatic forecasting system and does not replace data-quality governance: if the sample is too small, data quality is unreliable, confounders cannot be controlled, or the goal is to prove a predetermined conclusion, escalate to human judgment.