Login Create Account

Conjoint Study Design Generator

Design Tool Analysis Tool →

Create optimized fractional factorial designs for choice-based conjoint (CBC) studies. Define attributes and levels, specify constraints, and generate balanced experimental designs ready for data collection.

OVERVIEW & WHAT THIS TOOL SOLVES

Before you can analyze a Choice-Based Conjoint (CBC) survey, you must decide which product combinations respondents will actually see. Testing 5 attributes with 4 levels each produces 1,024 possible product profiles (45)—far too many for any human to evaluate. Conjoint Study Design Generator builds optimized fractional factorial choice designs that maximize statistical D-efficiency, level balance, and attribute orthogonality so you can estimate every feature's part-worth utility from just 8–15 choice tasks per respondent.

🎯 Questions It Answers

  • How do we compress hundreds of possible product combinations into a compact, unbiased set of survey choice tasks?
  • Does our experimental design show every attribute level equally often (balance) and independently of other attributes (orthogonality)?
  • How many respondents do we need to recruit so every part-worth parameter is estimated with reliable statistical power?

📂 What Data You Need

  • 3 to 8 Product Attributes: Core features under consideration (e.g., Brand, Price, Storage, Battery Life)
  • 2 to 6 Levels per Attribute: Specific values or price points to test for each attribute
  • Task Structure & Prohibitions: Tasks per respondent (8–15), alternatives per task (3–4), optional "None" option, and any forbidden level pairs

💡 What You'll Discover

  • Design Quality Scorecard: Profile Information Index (0–100%), Level Balance score, and Orthogonality diagnostics
  • Minimum Sample Size Guidance: Recommended respondent count based on Orme's rule of thumb and parameter degrees of freedom
  • Survey-Ready Exports: Downloadable CBC design matrix, respondent simulator dataset, and Qualtrics-compatible layout
🔍 Key Nuances & Common Traps to Watch For
  • Beware of Over-Using Prohibited Pairs: Forbidding certain combinations (e.g., prohibiting the "flagship brand" from appearing at the "lowest price") feels realistic, but heavy constraints destroy orthogonality—making it statistically impossible to separate brand utility from price sensitivity!
  • Number of Levels Bias: Respondents naturally pay more attention to attributes that span more levels (e.g., testing 6 price points alongside 2 warranty levels). Try to keep the number of levels reasonably balanced across attributes (typically 3–5 levels each).
  • Survey Fatigue vs. Information per Respondent: Asking 12 tasks with 3 alternatives per task is usually the sweet spot. Pushing past 15–18 choice tasks causes respondents to simplify or click randomly, degrading data quality.
📐 Under the Hood: Model Equations & Variable Definitions

1. Full Factorial Size, Parameter Degrees of Freedom & D-Efficiency

\[ N_{\text{full}} = \prod_{k=1}^{K} L_k, \qquad p = \sum_{k=1}^{K} (L_k - 1), \qquad D\text{-eff} = \left| \frac{1}{N_{\text{profiles}}} \mathbf{X}^{\top}\mathbf{X} \right|^{1/p} \]

💬 Read it in plain English: The full factorial multiplies the number of levels across all attributes, p counts the independent part-worth parameters to estimate, and D-efficiency measures how tightly the chosen subset of profiles packs statistical information into the design matrix X compared to an ideal orthogonal design.

Nfull
Full Factorial UniverseOUTPUT ⓘ
Total number of unique product profiles possible across all K attributes (∏ L_k).
K, Lk
Attributes & Levels per AttributeYOUR DATA ⓘ
Total number of attributes K and number of levels L_k defined for attribute k.
p
Main-Effect Parameters (df)OUTPUT ⓘ
Total non-reference dummy parameters to estimate across all attributes: ∑ (L_k − 1).
D-eff, X′X
D-Efficiency & Information MatrixINFERENCE ⓘ
Geometric mean of the eigenvalues of the normalized effects-coded information matrix (1/N)·X′X.

2. Orme's Minimum Sample Size Rule of Thumb for CBC

\[ N_{\text{min}} \ge \frac{500 \cdot L_{\max}}{T \cdot A} \]

💬 Read it in plain English: To estimate every attribute level's part-worth reliably, your total sample size times tasks per person times non-None alternatives per task should give every level of your largest attribute at least 500 appearances across the study.

Nmin
Minimum Recommended Sample SizeOUTPUT ⓘ
Minimum number of survey respondents needed for stable aggregate main-effect utilities.
Lmax
Largest Attribute Level CountYOUR DATA ⓘ
Maximum number of levels L_k in any single attribute (for main effects).
T
Choice Tasks per RespondentYOUR DATA ⓘ
Number of choice screens each respondent completes (typically 8 to 15).
A
Product Alternatives per TaskYOUR DATA ⓘ
Number of product concepts shown side-by-side per task (excluding the "None" option).

CHOOSE YOUR STARTING POINT

📚

Start from a Case Study

Load a pre-configured study to see how the generator works, then customize it.

🛠️

Build from Scratch

Define your own product attributes and levels from a blank canvas.