What kinds of questions trigger Category Comparison?

Two kinds. First, questions asking whether two categorical fields are related, such as is region associated with churn status, or does weather condition vary by city. Second, questions asking whether a category mix shifted between two periods, such as did the customer mix shift toward enterprise, or did the proportion of mobile orders change.

How is this different from asking AskEdi to compare a numeric metric across a category, like revenue by region?

Comparing a numeric metric across a category, such as why did revenue drop in the West region, is answered by Root Cause Analysis, because a numeric metric is being explained. Category Comparison is for questions with no numeric metric at all, just whether two categorical fields relate to each other or whether a category's mix changed.

For a composition shift question, how does AskEdi know which period is earlier?

AskEdi determines the two periods directly from the meaning of your question, not by sorting the period labels alphabetically. This matters because a label like Q3 2024 sorts alphabetically after Q1 2025 despite being chronologically earlier, so AskEdi resolves the actual chronological order rather than relying on text sorting.

What if the relationship is statistically significant but the effect is very small?

AskEdi states both the significance and the effect size together, and frames the strength honestly as weak, moderate, or strong. A statistically significant but weak association is reported as exactly that, not inflated into a major finding.

Does Category Comparison work across languages other than English?

Yes. Classification is based on the meaning of your question, not English keyword matching, so you can ask in any of AskEdi's supported languages.

Category Comparison | Edilitics Docs

How AskEdi tests whether two categories are really related, or whether a category mix genuinely shifted, using real statistical tests instead of a visual read.

Category Comparison is AskEdi's response to questions that test a relationship between two categorical fields, or whether a category's mix genuinely changed between two periods. Neither question has a numeric metric to explain, so AskEdi runs a dedicated statistical test rather than the regression and contribution math used for a "why did this number move" question.

When Category Comparison Activates

AskEdi classifies the intent of your question automatically. There are two distinct shapes, both routed to Category Comparison.

Asks whether two categorical fields are related to each other:

"Is shipment direction associated with on time delivery outcome?"
"Does weather condition vary by city?"
"Is region associated with churn status?"

Asks whether a category's mix changed between two periods:

"Did the customer mix shift toward enterprise?"
"Did the proportion of mobile orders change?"

A question that mentions a category only as a filter on a numeric metric, for example "why did revenue drop in the West region," is answered by Root Cause Analysis instead, the metric being explained decides the response type, not whether a category is mentioned.

What Data This Needs

Category Comparison needs genuinely categorical columns, not just any column with repeated values:

A small number of distinct values on each column. Both columns in an association question need to be truly categorical, region, status, or type, not a high-cardinality field like an order ID that happens to repeat.
At least two values in each category. A column already filtered down to a single value has nothing to test against.
A usable date column, for a composition shift question. Comparing a category's mix between two periods needs a real date or timestamp column with enough range to define two distinct periods to compare.

AskEdi checks that both categorical columns genuinely have a small number of distinct values before answering, and for a composition shift question, that there is a usable date column with two clear periods to compare. If either check fails, AskEdi explains why and suggests what to ask instead. No credit is consumed for that response.

What a Category Comparison Response Includes

Every Category Comparison response is four sections, shorter than Root Cause Analysis because it tests exactly one relationship, association or shift, never both at once.

1. Observed Pattern

States what is being compared, the raw shape of the data, and why the comparison matters in plain terms.

2. Statistical Test Result

The real computed result, never a visual estimate.

States whether the two categories are statistically associated, the significance result, and the effect size together, never significance alone. Effect size is described as weak, moderate, or strong.

With enough data, even a very weak relationship can be statistically significant. AskEdi always cites the effect size alongside significance, so a real-but-tiny association is not overstated as a meaningful finding.

Renders as a table, one row per category value, comparing the share each category held in the earlier period against the later period:

Category	Earlier period	Later period	Change	Significant

3. Category Breakdown

Names the specific category values driving the result, not just the aggregate statistic. For an association question, this describes which combination of categories appears over or under represented. For a composition shift question, this highlights the category with the largest real change.

4. Decision & Impact Summary

The closing section renders as a highlighted card, not plain text. It contains three labeled lines:

Line	What it contains
Recommended action	A concrete next step grounded in the specific category values identified in the breakdown above.
Evidence	The real test result and effect size (association) or the real percentage-point shift (composition) that supports the recommendation.
Confidence	High, medium, or low. Cites the real significance and effect-size result, the AI does not grade its own confidence.

What This Looks Like

A shortened example, for a question like "is shipment direction associated with on time delivery outcome?":

Observed Pattern This analysis compares shipment direction in 2 distinct categories against on time delivery outcome in 2 distinct categories.

Statistical Test Result Yes, there is a statistically significant association. A p-value of less than 0.001 means this result is very unlikely to have occurred by chance. The effect size of 0.25 falls in the moderate range.

Category Breakdown One shipment direction is systematically more concentrated in on time outcomes than the other, though both directions still contain a mix of results.

Decision & Impact Summary (highlighted card) Recommended action: Manage forward and return shipments as separate service-performance cohorts. Confidence: medium.

The real response is longer and grounded in your own column names and numbers. This shows the shape, not the content.

Verifying the Numbers Behind the Answer

The contingency table, the significance test, and the effect size are all computed server-side before the response is written. To inspect the work directly:

Analysis view: shows the exact query that ran against your data source.
Methodology Notes: shows the real test statistic, the sample size, and the effect size with a plain-language strength label, independent of what the response text says.

Category Comparison