Administer, score and record validated personality and affect questionnaires.

What this is

Collecting validated self-reports is the unglamorous half of an affective-computing pipeline. This package administers published psychometric instruments, scores them correctly, and records every response with the provenance needed to reproduce the score months later.

It is built around one idea: an instrument is data, not code. Items, response ranges, subscale membership and reverse keys are declared as values; a single vectorised scorer turns responses into scores without knowing which questionnaire it is holding. Adding an instrument adds no arithmetic.

Instruments

_{BFI-2 and BFI-2-XS: Soto & John (2017). BFI-10: Rammstedt & John (2007). VAS-F: Lee, Hicks & Nino-Murcia (1991). See docs/instruments.md for full citations and licence notes.}

Quickstart

Score responses you already have:

pip install personality_questionnaire

import personality_questionnaire as pq

result = pq.score(pq.get("bfi2"), answers)  # answers: (n_participants, 60)
result.by_level("domain")  # {"openness": array([...]), ...}
result.as_dict()  # one participant, every subscale

Administer one at the terminal:

pq list                                  # what is available
pq info bfi2                             # items, subscales, citation
pq run bfi2 --participant P01            # ask the questions, score the answers
pq score bfi2 --input answers.csv        # score a file

How it works

flowchart LR
    I["instruments/<br/><i>pure data</i>"] --> R[registry]
    R --> S["scoring<br/><i>one vectorised scorer</i>"]
    S --> C[cli]
    S --> D[db]
    S --> U[ui]

Design decisions

Instruments are Python literals, not data files. A literal is checked by the type checker, validated at import, and present in the wheel by construction. A shipped CSV is checked by nothing until a participant has already answered every item — and the two scale files this repo used to carry were never read by any code path and misspelled neuroticism, which is exactly how unread data drifts.

Reverse-keying belongs to the subscale, not the item. The VAS-F scores its five energy items forward in Energy and reversed in Fatigue (composite), so a per-item mask cannot express both. The reflection is folded into a signed weight matrix, which also means every subscale at every level is computed by one matrix multiplication.

Both hierarchy levels are declared flat. The BFI-2's five domains and fifteen facets are siblings, each listing its own item numbers, rather than domains being composed from facets. The arithmetic is identical and the flat form scores both levels in a single pass.

Polarity is recorded as data. Every subscale carries a higher_is string, so no consumer has to infer direction from a name — the inference that produced the bug below.

Development

make dev          # install everything
make fix          # format and autofix
make check        # lint, type-check, test, docs -- mirrors CI
make check-ci     # the same, in a throwaway venv built like CI's

Tests are unittest under coverage, mirroring the package layout in tests/.

Related work

PersonalityLinMulT predicts perceived Big Five traits from video. This package sits on the other side of that problem: it collects self-reported ground truth, on the same [0, 1] scale and in the same openness, conscientiousness, extraversion, agreeableness, neuroticism column order, so an exported BFI-2 record drops into a self-report-versus-perception comparison. The two are deliberately uncoupled in code — this package has no ML dependencies.

Citation

@software{fodor_personality_questionnaire,
  author = {Fodor, Ádám},
  title  = {personality_questionnaire: administering and scoring validated psychometric instruments},
  url    = {https://github.com/fodorad/personality_questionnaire},
}

Contact

• Ádám Fodor (fodorad201@gmail.com) — adamfodor.com