Cecilia-lll/IPIP-BFFM-openpsy

Dataset Summary

This dataset contains human questionnaire responses (IPIP Big Five format) plus a reproducible quality-control (QC) pipeline and derived outputs.

The original file is processed to identify low-quality responses using technical, response-pattern, multivariate, and context-based checks.
Two main analysis files are produced:

• a fully flagged version (all rows kept, QC flags added)
• a cleaned version (rows excluded if any exclusion flag is triggered)

---

Source Data

• Original input: data-IPIP-BFFM-openpsy.csv (tab-separated in the processing notebook)
• Items include 50 IPIP Big Five items:
  • traits: EXT, EST, AGR, CSN, OPN
  • item columns: <TRAIT>1 to <TRAIT>10
• Timing columns: <TRAIT><i>_E (milliseconds)
• Additional context columns used in QC:
  • IPC: the count of records from the same IP
  • screenw: The width of the user's screen in pixels
  • screenh: The height of the user's screen in pixels
  • testelapse: The time in seconds spent on the page with the survey questions

More descriptions of source data: codebook.txt

---

Processing and Quality Control

Processing is implemented in: data_process.ipynb.

QC checks applied

A row is flagged by the following checks.

These checks are designed to remove responses that are likely invalid for psychometric analysis.
In plain terms, we flag records that look duplicated, incomplete, implausibly fast, overly repetitive, statistically unusual, or inconsistent with expected device/session context.

1. Technical checks

   Why this method: Technical checks catch data integrity problems before behavioral interpretation.
   Why these thresholds: The thresholds follow direct validity constraints from the questionnaire format (not statistical tuning).

   • flag_ipc_duplicate: IPC != 1
IPC is the count of records from the same IP in this dataset. A value of 1 means the response is unique by IP; values >1 indicate possible duplicate participation.
   • flag_missing_items: any missing value in 50 item responses
     Complete item data are required for scale scoring and multivariate QC.
   • flag_item_out_of_range: any item outside allowed range [1, 5]
     IPIP items are 1–5 Likert responses; values outside this range are invalid entries.

2. Engagement / response-pattern checks

   Why this method: Very fast or highly repetitive patterns are common indicators of low-effort responding.
   Why these thresholds: Cutoffs are conservative heuristics intended to detect clear low-quality patterns while retaining normal variation.

   • flag_speeding: mean item time < 1.5 seconds
     On a 5-point personality item, average times below 1.5s are often too short for careful reading/selection.
   • flag_straightlining: unique response count across items <= 2
     Using only 1–2 response values across 50 items suggests minimal differentiation.
   • flag_longstring: longest run of identical consecutive responses >= 10
     Long uninterrupted runs of the same answer indicate possible satisficing/automatic responding.
     Computed in questionnaire presentation order: EXT1, AGR1, CSN1, EST1, OPN1, EXT2, ....

3. Multivariate outlier check

   Why this method:
   We treat each respondent as a 50-dimensional feature vector (one value per questionnaire item), then standardize items using z-scores so all items are on a comparable scale. Mahalanobis distance then measures how far each respondent is from the average response pattern, while accounting for correlations among items (so related items are not double-counted as independent signals).

   Why this threshold:
   Each respondent gets a Mahalanobis distance value (D^2). We compare (D^2) to a theoretical chi-square distribution with 50 degrees of freedom (matching the 50 item features). We use a strict cutoff, alpha = 0.001, which means only responses in roughly the most extreme 0.1% of expected values are flagged. In practice, this targets only highly atypical multivariate response patterns that are unlikely under normal responding behavior.

   • flag_mahalanobis: Mahalanobis (D^2) on z-scored complete item responses
   • threshold: chi-square upper-tail cutoff with alpha = 0.001 and df = 50

4. Context-variable outlier checks

   Why this method: Device/session variables can reveal implausible records (e.g., extreme screen sizes or elapsed times).
   Why this threshold: The IQR rule (1.5 × IQR) is a standard robust outlier rule that does not assume normality.

   • IQR rule with multiplier 1.5 for:
     • screenw: The width the of user's screen in pixels
     • screenh: The height of the user's screen in pixels
     • testelapse: The time in seconds spent on the page with the survey questions
   • flags:
     • flag_screenw_outlier
     • flag_screenh_outlier
     • flag_testelapse_outlier

Exclusion rule

exclude_any = True if any of the exclusion flags is True:

• flag_ipc_duplicate
• flag_missing_items
• flag_item_out_of_range
• flag_speeding
• flag_straightlining
• flag_longstring
• flag_mahalanobis
• flag_screenw_outlier
• flag_screenh_outlier
• flag_testelapse_outlier

---

Files in this folder

• data-IPIP-BFFM-openpsy.csv
  Original raw survey export.

• data-flagged-qc.csv
  Full dataset with all derived QC metrics and flag columns.

• data-cleaned.csv
  Subset where exclude_any == False.

• qc-flag-summary.csv
  Per-flag counts and proportions.

• qc-flag-overlap.csv
  Correlation matrix of exclusion flags (flag overlap structure).

• data_process.ipynb
  End-to-end processing notebook used to generate all derived files.

---

Intended Use

• Human response baseline data for comparison studies (e.g., LLM vs human response characteristics).

---

Limitations and Notes

• Exclusion thresholds are rule-based and may affect sample composition.
• qc-flag-overlap.csv reports correlations between binary flags (not causal relationships).
• If reused in another setting, verify assumptions behind timing thresholds and IQR-based context filtering.

---

Reproducibility

Run data_process.ipynb to reproduce:

• data-flagged-qc.csv
• data-cleaned.csv
• qc-flag-summary.csv
• qc-flag-overlap.csv