odyn-network/quant-data-example

Quant Data Example

De-identified responses from a structured US consumer survey on travel planning, multi-attraction pass preferences, lifestyle interests, aesthetic taste, values and demographics. 2,007 respondents, 394 survey variables (33 multi-item blocks and 36 single-choice variables), fielded 2022-03-22 to 2022-03-28.

• In responses and wide, one row = one respondent; in everything, one row = one answer. Every file carries the same opaque respondent_id (r0001...), assigned after a seeded shuffle; it is not derived from any source identifier.
• Most variables are categorical choices or 'select all that apply' items, so the data suits segmentation, preference modelling and propensity experiments, and demonstrating survey analytics without personal data.
• Self-reported panel data from one week, US only. No behavioural outcomes are observed.

Quick start

from datasets import load_dataset
responses = load_dataset("odyn-network/quant-data-example", "responses", split="train")  # nested: section > block > answers, 1 row per respondent
questions = load_dataset("odyn-network/quant-data-example", "questions", split="train")  # tree as a flat table: node_type, parent_id
everything = load_dataset("odyn-network/quant-data-example", "everything", split="train")  # flat: 1 row per answered item
wide = load_dataset("odyn-network/quant-data-example", "wide", split="train")  # 1 row per respondent, for modelling
wide.to_pandas()["age_band"].value_counts()

import pandas as pd
df = pd.read_parquet("hf://datasets/odyn-network/quant-data-example/data/respondents.parquet")

Files

File	What it is
data/responses.parquet / data/responses.jsonl	Start here. One nested record per respondent mirroring the question hierarchy: sections > blocks > answers. Each block carries the page and this respondent's seconds on it; each answer carries the question id, item label (omitted when it equals the response) and response. The Parquet is what the Hub loads and shows as nested JSON; the .jsonl holds identical content as plain JSON lines for agents and scripts.
data/questions.json	The question tree with metadata: sections > blocks (parent questions) > items, with answer options, answer counts, % answered, page and page timing (median, quartiles), plus a file_guide explaining every field of every file.
data/questions.parquet	The same tree as a flat table: one row per section / block / item with node_type and parent_id.
data/everything.parquet	Flat alternative: one row per answered item (228,175 rows) with question labels, page and seconds. Handy for SQL / pandas group-bys.
data/respondents.parquet	For modelling. Wide, 394 readable snake_case columns + respondent_id + total_seconds. Select-all items are nullable booleans (True = selected, null = not selected or not shown); scales are int8; everything else is a string.
input.jsonl	Same data, original schema and original column names (Item:SOURCE_CODE), HTML removed, empty answers omitted (a missing key means no answer), plus respondent_id.
input_structured.jsonl	Same data in the original nested layout {"meta": {...}, "responses": {block: {item: value}}}. See Fixes to the structured file.
original_questions.json	Column-by-column mapping to the source instrument, same schema as before with added role, released, duplicate_of, col_id. Includes the dropped columns (names only).
docs/codebook.md / docs/codebook.csv	Full data dictionary: every variable, option, answer count and distribution. Give this file to an LLM or agent along with the Parquet.
docs/dropped_columns.csv	Every removed source column, why, and how many respondents had a value (counts only).
docs/example_respondents.md	Four respondents rendered as question -> answer.
docs/stats.json	Machine-readable summary statistics used in this card.
scripts/build_release.py	Deterministic script that produced everything here from the raw export.

Contents at a glance

Section	Blocks	Variables
Multi-attraction passes: awareness, consideration, attitudes	12	76
Travel behaviour and intent	7	50
Lifestyle, interests and brands	10	161
Media habits	1	6
Aesthetic and style preferences	11	21
Values, motivations and psychographics	9	60
Demographics	9	9
Socio-economic	9	9
Screener / attention items	1	2

Respondent profile (share of 2,007):

Age band	n	%
18-24	90	4.5%
25-34	361	18.0%
35-44	320	15.9%
45-54	357	17.8%
55-64	393	19.6%
65 and above	486	24.2%

Gender	n	%
Female	1,174	58.5%
Male	833	41.5%

Generation	n	%
Baby boomers	700	34.9%
Generation X	585	29.1%
Millennials	533	26.6%
Generation Z	100	5.0%
The Silent Generation	89	4.4%

Sample skews older: 43.8% are 55 or over, and 80.4% report their ethnicity as White. Weight or stratify before generalising.

A few headline results

Agreement with statements about multi-attraction passes (Agree + Strongly agree, % of those who answered):

Statement	% agree
They are a great idea	80.1%
They are a convenient option for me when planning my trip	77.6%
They provide lots of activities/excursions to choose from	77.5%
They allow me to make the most from my trip	75.7%
They are an exciting choice to have as part of my trip	72.4%
They provide an easy one-stop-shop for everything I want to see	72.2%
They offer excellent value for money	70.3%
They make me feel confident when planning my trip	68.7%
They are easy to understand	66.3%

Preferred destination types (select all; % of the 2,007 respondents who answered):

Type	n	%
Beach or seaside resort	1,045	52.1%
City or Metropolis	859	42.8%
Rural or countryside locales	453	22.6%
Mountain or alpine resort	354	17.6%
Something else	268	13.4%
Culturally/Spiritually significant place	192	9.6%
Niche or extreme destinations (e.g. Antarctica)	51	2.5%

Travel frequency:

Frequency	n	%
Once a year or about once a year	768	38.3%
Multiple times a year	692	34.5%
Less than once a year	547	27.3%

First-choice pass (answered by 755 respondents; pass brands are shown as letters):

Pass	n	%
Pass B	254	33.6%
Pass A	222	29.4%
Pass F	94	12.5%
Pass C	61	8.1%
Pass D	46	6.1%
Pass E	42	5.6%
Pass I	18	2.4%
Pass H	18	2.4%

How to read missing values (important)

Many blocks were shown to subsets of respondents (for example the Pass A-K blocks were answered by 816 of 2,007). In select-all blocks the export stores a value only when an option was ticked, so null means 'not ticked' or 'not shown' and the two cannot be told apart. For rates, use as denominator the respondents who answered at least one item in the block; docs/codebook.md gives both denominators for every item. Overall 28.9% of respondent x variable cells are answered.

Structure: parents and children

Questions are nested. A section (theme) contains blocks, and a block is a parent question such as a select-all list or a rating grid; its items are the child rows or options. A single-choice question is a block with exactly one item. 9 sections, 69 blocks, 394 items. data/responses.jsonl stores answers in the same nesting, and data/questions.json describes it.

A block in questions.json (items truncated):

{
 "section_id": "pass_funnel",
 "label": "Multi-attraction passes: awareness, consideration, attitudes",
 "blocks": [
  {
   "block_id": "city_trip_activities",
   "label": "Activities on a city trip (select all)",
   "answer_type": "select_all",
   "question_wording": null,
   "page": 4,
   "page_presentation": 1,
   "page_seconds": {
    "median": 22,
    "p25": 17,
    "p75": 32,
    "n_viewers": 2007
   },
   "n_answered": 1934,
   "pct_answered_of_all": 96.4,
   "options": [
    "Have a stroll or cycle around town",
    "Buy a souvenir",
    "Find the best views"
   ],
   "options_ordered": false,
   "source_code": "AIDA_WW_CBA_CB_27092021",
   "items": [
    {
     "question_id": "city_trip_activities__have_a_stroll_or_cycle_around_town",
     "label": "Have a stroll or cycle around town",
     "n_answered": 1005,
     "pct_of_block_respondents": 52.0,
     "source_column": "Have a stroll or cycle around town:AIDA_WW_CBA_CB_27092021"
    },
    {
     "question_id": "city_trip_activities__buy_a_souvenir",
     "label": "Buy a souvenir",
     "n_answered": 1264,
     "pct_of_block_respondents": 65.4,
     "source_column": "Buy a souvenir:AIDA_WW_CBA_CB_27092021"
    }
   ]
  }
 ]
}

A respondent in responses.jsonl (truncated):

{
 "respondent_id": "r0001",
 "total_seconds": 2318,
 "sections": [
  {
   "section_id": "travel_behaviour",
   "label": "Travel behaviour and intent",
   "blocks": [
    {
     "block_id": "destination_type",
     "label": "Preferred destination types (select all)",
     "page": 20,
     "page_seconds": 7,
     "answers": [
      {
       "question_id": "destination_type__beach_or_seaside_resort",
       "response": "Beach or seaside resort"
      },
      {
       "question_id": "destination_type__mountain_or_alpine_resort",
       "response": "Mountain or alpine resort"
      }
     ]
    }
   ]
  },
  "..."
 ]
}

Timing

The survey platform logged how long each respondent spent on each page. Questions on a page share that time, so time is per page, not per question. In the export a page's time column follows the questions on that page; that rule maps 54 pages to 394 of 394 variables, and was checked against the data: on 40 of 54 pages the respondents with a time are exactly those who answered something on the page, and on the rest no respondent answered a page without having a time (0 violations).

• page_seconds (in responses and everything) is the respondent's time on that page; total_seconds (in respondents) is their total time. Units are not documented in the source; the typical total of about 1,164 suggests seconds.
• Times are capped at the 99th percentile of each page (total at 6,241) because the raw maxima are idle sessions (hours or days). Use the median and quartiles in questions for typical timing.
• Timing is not in input.jsonl or input_structured.jsonl, which keep the original schema.
• Individual timing patterns are a possible link to the source platform's logs. Do not use them to try to match respondents to other records.

Column naming

<block_id>__<option> for grid items (e.g. destination_type__beach_or_seaside_resort), and a plain name for single-choice variables (e.g. age_band). The docs/codebook.csv column original_column maps every id back to the source export, and input.jsonl keeps the source column names. Block titles are curated from the option wording because the source instrument only provides opaque codes (for example SBEH_WW_DESTTYPE_CB_07062021). Only 4 variables/blocks have question wording in the source; those are marked in the codebook, and nothing was invented for the rest.

What was removed, and what residual risk remains

536 columns in the raw export, 394 released, 142 removed:

Removed category	Columns
telemetry	66
empty	51
constant	6
identifier	5
webhook	5
open text	3
timestamp	2
comments	2
location geo	2

• Identifiers: response / contact / session / respondent UUID and the row index. Replaced by an opaque id after a seeded row shuffle (so file order carries no timing information).
• Exact timestamps (only the fielding window above is reported) and geo-IP city / region.
• Page timing and total-time telemetry, webhook / redirect plumbing.
• Pass brand names anonymised. In two questions the source showed real brand names for four of the passes; they were replaced by the letter used for the same pass elsewhere in the survey, so each pass now has a single consistent label (A-K). The key is held by the dataset owner. Generic category lists (car makers, tech brands, booking channels such as Expedia or Groupon) are unchanged.
• Free-text answers (3 columns: two brand-awareness text boxes and the 'something else' write-in for causes) are withheld because they may contain personal details.
• Columns that are constant for every respondent (completion status, language, country, opening agreement item, quota, one screener item) are not repeated; all respondents are marked Complete and English; Country is 'United States' for every respondent who has a value (2,006 of 2,007); everyone answered 'Yes, I agree' to the opening agreement item (S01Q01); the screener item ATR_01 was 'No' for all.
• Columns that are empty for every respondent in this export.

No names, contact details, free text, timestamps, or location finer than the respondent's declared state of residence remain. Residual re-identification risk is not zero. The survey records many demographic attributes at once, so respondents can be distinguishable by combination:

Attributes combined	Respondents unique in the file	Smallest group
age band + gender + declared state	123 of 2,007 (6.1%)	1
+ ethnicity	371 of 2,007 (18.5%)	1
12 demographic variables together	1,848 of 2,007 (92.1%)	1

Someone who already knows many facts about a specific respondent could locate their row. Do not attempt to re-identify respondents, and do not link this file to other datasets for that purpose.

Known issues in the source data (kept as-is, not recoded)

• Empty blocks. The social-media, TV-channel and news-outlet blocks (social_media_use, tv_channels, news_outlets, 41 columns) and Pass G-K in the awareness block contain no responses in this export, so they are not released. Media habits are represented only by the six-channel ranking (media_rank); its direction (is 1 most or least used?) is not documented in the source.
• Label inconsistencies. The NPS group appears as both Pasives (n=77) and Passives (n=8); Saftety and Fariness are misspellings in the source; the same option can appear with a straight or a curly apostrophe (It's vs It’s). Recode before analysis if you want them merged.
• Two presentations of one list. The 'life priorities' block (10 options) appears twice in the source export (second set suffixed .1; the two sets differ for many respondents). Both are released (life_priorities__* and life_priorities_p2__*); what the second presentation represents is not documented. The only other .1 column with content was a free-text write-in and is withheld.
• Partly unordered scales. Age, income bands, income satisfaction, agreement and consideration scales have an explicit order in the codebook. Values such as 'Prefer not to say' are listed as unordered.
• Self-reported, single-week, US-only, older-skewed panel sample; response bias is likely; agreement with every pass statement is between 66.3% and 80.1%, which can reflect acquiescence.

Fixes to the structured file

Relative to the previous input_structured.jsonl: values are now always strings (previously a list in 2,046 places, i.e. a string-or-list mix that strict typed readers reject, and which also embedded free-text write-ins); the second presentation is a separate block key ending .1 instead of being merged; HTML markup is removed from keys and values; identifiers, free text and constants are removed; respondent_id is added to meta. Checked against the previous structured file: 226,331 compared cells, differences: none.

Reproducibility

python scripts/build_release.py --raw <raw dir> --out <release dir> reproduces every file from the raw export (seed 20261001). The build asserts that wide, long, flat and structured outputs contain the same 228,175 answered cells and that no identifier-like pattern (UUID, session id, timestamp, URL fragment) appears in the released JSONL.

Intended use and limitations

Intended for demonstrating and testing survey-analytics methods (descriptive summaries, segmentation, factor analysis, propensity modelling) on non-personal data. Not suitable for inferring behaviour, for population estimates without weighting, or for any attempt to identify individuals. For modelling, validate out of sample.