Declare a choice-experiment (conjoint) design

Generates and stores a conjoint profile set and task schedule as part of the instrument's pre-declared contract. This is a design generator, not an estimator: it fixes what respondents will be shown, and it does not fit or analyse choice models.

Usage

sf_conjoint_design(
  id,
  attributes,
  method = c("full", "balanced", "random"),
  n_profiles = NULL,
  n_alternatives = 2L,
  n_tasks = NULL,
  blocks = 1L,
  seed = NULL,
  profiles = NULL,
  label = NULL
)

Arguments

id: Design identifier. Must start with a letter and contain only letters, numbers, and _ characters.
attributes: Named list of character vectors, one per attribute, giving that attribute's levels. At least 2 attributes, each with at least 2 levels.
method: One of "full", "balanced", or "random". Ignored when profiles is supplied.
n_profiles: Number of profiles to keep. Required for "balanced" and "random", ignored for "full".
n_alternatives: Alternatives shown per choice task, default 2.
n_tasks: Choice tasks per block. Defaults to as many whole tasks as the profile set supports.
blocks: Number of blocks the tasks are split across, default 1.
seed: Integer seed. Generated and stored when not supplied.
profiles: Optional data frame of pre-built profiles, one column per attribute. Supplying this bypasses generation and declares the design as given.
label: Human-readable label.

Value

An object of class sf_conjoint_design with $profiles, $tasks (long format, one row per block, task, and alternative, ready for a choice model), $balance, and the declaration that produced them.

Details

The design is reproducible by construction. seed is always recorded, and generated when not supplied, so regenerating from the stored declaration returns the identical profiles and tasks.

Choosing a method

"full" enumerates every combination, which is exact but grows fast: 4 attributes at 3 levels each is 81 profiles. "random" takes a seeded random subset of that size. "balanced" samples repeatedly and keeps the subset with the most even level spread and the weakest association between attributes.

"balanced" is a search, not a construction. It does not produce a catalogued orthogonal fractional factorial and does not claim the guarantees of one. The achieved balance is reported in $balance so the design can be inspected rather than trusted. A study needing a specific D-optimal or orthogonal design should generate it elsewhere and pass it in through profiles, which keeps the declaration in the contract either way.

Examples

design <- sf_conjoint_design(
  "hotel_dce",
  attributes = list(
    price    = c("50", "100", "150"),
    board    = c("room only", "breakfast"),
    distance = c("beachfront", "10 min walk")
  ),
  method = "balanced", n_profiles = 6, n_alternatives = 2, seed = 42
)
design$profiles
#>   profile_id price     board   distance
#> 1         p1    50 breakfast beachfront
#> 2         p2   100 room only beachfront
#> 3         p3   150 breakfast beachfront
#> 4         p4   150 room only beachfront
#> 5         p5    50 room only beachfront
#> 6         p6   100 breakfast beachfront
design$tasks
#>   block task alternative profile_id
#> 1     1    1           1         p2
#> 2     1    1           2         p4
#> 3     1    2           1         p3
#> 4     1    2           2         p6
#> 5     1    3           1         p5
#> 6     1    3           2         p1