toolEnglish

Columns and Quality Review

How to inspect per-column type, missingness, uniqueness, statistics, top values, descriptor notes, and review outcomes.

EnglishEspaƱolFranƧais

Documentation follows the current system language. Missing translations fall back to the page's canonical language.

Columns And Quality Review

The Columns module is the detailed review surface. It shows one row per profiled column and supports searching by column name and filtering by inferred type.

What To Read First

Start with columns that have:

  • high missing percentage
  • unexpectedly high or low unique count
  • quality flags
  • identifier-like behavior
  • statistics that conflict with domain expectations
  • top values that reveal inconsistent labels

Column Fields

Each row contains:

  • Column: original field name plus descriptor notes when available
  • Type: inferred type used to choose statistics
  • Missing %: proportion of rows without usable value
  • Unique: number of distinct values
  • Stats: type-specific statistics
  • Top Values: most frequent values for categorical or text-like fields

Numeric Columns

For numeric fields, compare min, max, mean, standard deviation, median, and p95. Large gaps between mean and median may indicate skew. Extreme min or max values may indicate unit errors or data-entry issues.

Categorical Columns

For categorical fields, inspect top values and unique count. Look for duplicated meanings such as "NY", "New York", and "new york"; inconsistent capitalization; unexpected empty labels; or categories that should not exist in the decision frame.

Date Columns

For datetime fields, confirm the range matches the intended analysis period. A stale max date can indicate an outdated extract. A surprising min date can indicate historical data pulled into a current-period decision.

Identifiers And Text

Identifier columns are often useful for joins and deduplication but usually not useful as explanatory variables. Text fields may need separate text analysis before they can be used in a structured decision model.

Review Outcome

After reviewing columns, classify the dataset into one of three states:

  • Ready for descriptive use: issues are minor or documented
  • Requires cleaning: issues are fixable but must be handled before decisions
  • Not fit for this question: missingness, scope, definitions, or quality problems make conclusions unsafe