AgentSkillsCN

data-validator

验证数据完整性——检查缺失值模式、变量分布、处理组平衡、样本量一致性，并与论文中的论断进行交叉比对。生成报告，无需修改原始文件。

SKILL.md
--- frontmatter
name: data-validator
description: Validate data integrity — check missingness patterns, variable distributions, treatment balance, sample size consistency, and cross-reference against manuscript claims. Produces a report without editing files.
disable-model-invocation: true
argument-hint: "[dataset filename]"

Validate Data Integrity

Check the project's cleaned dataset for integrity issues: unexpected missingness, implausible distributions, treatment imbalance, and consistency with manuscript claims.

Steps

  1. •

    Read CLAUDE.md to find:

    • •Cleaned dataset location and filename
    • •Key variable names (treatment, outcomes, controls)
    • •Expected sample size
    • •Manuscript location (for cross-checking claims)
  2. •

    Identify dataset to validate:

    • •If $ARGUMENTS is a specific filename: validate that file
    • •Otherwise: validate the primary cleaned dataset listed in CLAUDE.md
  3. •

    Write and run a validation R script that checks:

    Basic structure:

    • •Number of rows and columns
    • •Variable names and types (numeric, character, factor)
    • •Duplicate row detection (by respondent ID if available)

    Missingness:

    • •Per-variable missing rate
    • •Flag variables with >20% missing
    • •Check if missingness is correlated with treatment assignment
    • •Identify rows with >50% missing values

    Treatment assignment:

    • •Frequency table of treatment variable
    • •Balance check: roughly equal N per arm
    • •Chi-squared test for uniform distribution

    Outcome variables:

    • •Range check (are indices bounded as expected?)
    • •Distribution summary (mean, SD, skewness)
    • •Correlation matrix among outcomes
    • •Check for floor/ceiling effects

    Control variables:

    • •Plausible ranges (e.g., age > 0 and < 120)
    • •Valid categories for factor variables (e.g., state codes)
    • •Flag any suspiciously uniform distributions

    Cross-dataset consistency (if multiple dataset versions exist):

    • •Compare row counts across variants
    • •Compare variable availability
    • •Check that subsetting is documented
  4. •

    Produce a validation report with:

    • •Dataset summary table (N rows, N cols, file size)
    • •Missingness heatmap description (top 20 variables by missing rate)
    • •Treatment balance table
    • •Outcome distribution summary
    • •List of flagged issues with severity
  5. •

    Save report to quality_reports/data_validation.md

  6. •

    IMPORTANT: Do NOT edit any data files or analysis scripts. Only produce the report. Issues are investigated after user review.

  7. •

    Present summary:

    • •Dataset dimensions
    • •Number of issues flagged by severity
    • •Most critical concerns highlighted