Skip to main content

Design a population

By default an experiment runs against a general population. Targeting narrows who answers.

Traits​

Traits are the characteristics a respondent can have. Fetch the catalogue before you target: the ids are what the API expects:

curl "$SUBCONSCIOUS_API/api/v1/traits" \
  -H "Authorization: Bearer $SUBCONSCIOUS_TOKEN"

Each trait carries a short and long description, a measurement type, and whether it is ordinal.

Target by trait​

Pass trait filters on the experiment request:

{
  "why_prompt": "What factors drive consumer choice of electric vehicles?",
  "population_traits": {
    "Travel frequency": ["Weekly"]
  }
}

Target by demographics​

US demographic targeting goes in target_population; outside the US, use non_us_target_population. Available axes include age, income, education, racial group, home ownership, household size, and households with children.

Validate before running​

Validation tells you whether a targeted population can actually be built. Over-constrained targeting is the most common cause of a disappointing run: each additional filter shrinks the pool, and a thin pool produces noisy estimates.

curl -X POST "$SUBCONSCIOUS_API/api/v1/populations/validate" \
  -H "Authorization: Bearer $SUBCONSCIOUS_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "age": [25, 45],
    "household_income": [50000, 150000],
    "state": "California",
    "gender": ["Female", "Male"],
    "number_of_records": 500
  }'

Use /api/v1/populations/validate-non-us for non-US populations.

If you need help defining the boundary, start a population recommendation from the research question, then poll the job. Treat the completed result as a proposal. Review its assumptions and warnings, then validate the proposed population before launch.

Population groups and external personas​

You can run against a named population group, or supply your own personas. These two are mutually exclusive: pick one. Supplying personas is the right move when you have first-party audience data; a named group is the right move when you want a familiar segment reproduced consistently across runs.

Practical advice​

  • Target on what plausibly changes the decision. Filters that do not affect the choice cost you precision and buy nothing.

Prince Charles and Ozzy Osbourne share every demographic attribute

Two people can match on every demographic filter and decide nothing alike. Target on what moves the choice.

  • Validate first, then run.
  • Keep the population fixed when comparing two designs. Changing the audience and the design at once makes the comparison meaningless.