Conjoint Analysis Sample Size Requirements
How Many Respondents Does a Conjoint Study Need?
The sample size for a conjoint analysis study depends on your design complexity, the number of attribute levels, and whether you need segment-level estimates. There's no universal answer, but there are reliable formulas and practical benchmarks that get you to the right number.
The short version: most standard CBC conjoint studies need 300-500 respondents. Simple designs can work with 200. Complex designs with segment analysis often require 500-1,000+. Below, we'll break down exactly how to calculate what your specific study needs.
The Sawtooth Rule of Thumb Formula
The most widely used sample size formula for choice-based conjoint (CBC) comes from Sawtooth Software. It calculates the minimum number of respondents needed to ensure each attribute level gets enough exposure across the dataset:
N >= 500 x c / (t x a)
Where:
- c = the largest number of levels in any single attribute
- t = number of choice tasks per respondent
- a = number of product profiles per task (excluding the "none" option)
Worked Example
Say you're running a study with these parameters:
- Largest attribute has 5 levels
- 12 choice tasks per respondent
- 4 profiles per task
N >= 500 x 5 / (12 x 4) = 2,500 / 48 = 52 respondents minimum
That's the mathematical floor. It means you have enough data for aggregate-level logit estimation. It doesn't mean you have enough for reliable individual-level modeling, segment analysis, or interaction effects.
Why 500 Isn't Enough
The formula targets 500 exposures per attribute level, which Sawtooth describes as a "bare minimum." In practice, 1,000 exposures per level is safer, which doubles the formula's output. The formula was also developed for aggregate estimation, not the hierarchical Bayesian (HB) models that are standard today. HB produces better individual-level estimates but benefits from larger samples to stabilize the population-level priors.
Practical Sample Size Benchmarks
| Study Type | Attributes | Levels per Attribute | Recommended Sample |
|---|---|---|---|
| Simple | 4-5 | 2-3 | 200-300 |
| Standard | 5-7 | 3-5 | 300-500 |
| Complex | 7+ | 4-5 | 500-1,000 |
| Segment analysis | Any | Any | 200-300 per segment |
These benchmarks assume CBC with 12-15 choice tasks and 3-4 profiles per task. If you're using fewer tasks or profiles, scale up the sample proportionally.
Factors That Increase Your Sample Needs
More Levels Per Attribute
The formula's numerator is driven by the attribute with the most levels. An attribute with 6 levels needs 20% more sample than one with 5, all else equal. If you can consolidate levels (e.g., combining "50 GB" and "100 GB" into "50-100 GB"), you'll reduce your sample requirement.
Segment-Level Analysis
This is the most common source of under-sampling. A 400-person study feels large, but if you want to compare preferences across three customer segments (enterprise, mid-market, SMB), each segment needs 200+ respondents for stable HB estimates. That's 600 minimum, not 400.
Plan your segments before you set your sample size, not after. Retrofitting segment analysis onto an undersized sample produces estimates with confidence intervals too wide to act on.
Interaction Effects
If you need to measure how two attributes interact (e.g., does the value of "unlimited users" change at different price levels?), you need more data than for main effects alone. Interaction models can require 50-100% more sample than main-effects-only models, depending on the number of interactions you want to estimate.
High "None" Selection Rate
When many respondents choose "none of these" across tasks, you're effectively losing data from those choices. If your pilot suggests a 40-50% none rate, increase your target sample by 30-50% to compensate.
How to Right-Size Without Overspending
Use the Formula as a Floor, Not a Target
Calculate the formula result, then multiply by 2-3x for practical reliability. A formula result of 52 means you should plan for at least 150-200, and 300+ if budget allows.
Increase Tasks Before Increasing Sample
Adding 3 more choice tasks per respondent is much cheaper than adding 100 more respondents. Going from 10 to 13 tasks gives you 30% more data per person. Just don't exceed 15-18 tasks or you'll trade sample efficiency for respondent fatigue.
Run a Power Simulation
The most precise approach: generate synthetic data matching your planned design, run the analysis, and check whether utility estimates are stable at your target sample size. This is the "random robots" method. Run the analysis at n=200, n=300, n=500, and see where standard errors drop below acceptable thresholds (typically 0.05 for main effects).
Pilot With 30-50 Respondents
A small pilot won't give you final results, but it will reveal problems: attribute levels that confuse respondents, a "none" rate that's too high, or completion times that suggest the survey is too long. Fix these issues before committing to full fielding.
Sample Size for Specific Methods
| Method | Minimum | Recommended | Notes |
|---|---|---|---|
| CBC (standard) | 200 | 300-500 | Most common; HB estimation |
| ACBC (adaptive) | 100-150 | 200-300 | Collects more data per person |
| Menu-Based | 300 | 400-600 | More parameters to estimate |
ACBC requires smaller samples than CBC because each respondent provides more information through the adaptive interview process. If you're sample-constrained, ACBC is worth considering for studies with 6+ attributes.
For precise estimates based on your design parameters, use the conjoint sample size calculator.
Frequently Asked Questions
Can I run a conjoint study with 100 respondents?
You can, but the results will be limited. At n=100, aggregate-level estimates (average preferences across all respondents) will be reasonable for simple designs. Individual-level estimates and segment comparisons won't be reliable. If budget constrains your sample to 100, keep the design simple (4-5 attributes, 2-3 levels) and don't attempt subgroup analysis.
Does doubling the sample size double the precision?
No. Precision improves with the square root of sample size. Doubling from 200 to 400 improves precision by about 41%, not 100%. You get diminishing returns past n=400-500 for most designs, which is why the 300-500 range represents the best cost-to-precision trade-off.
How many respondents per segment do I need?
Plan for 200-300 per segment. With fewer than 150 per segment, HB estimates become unstable and confidence intervals widen to the point where you can't distinguish between segments reliably.
Related Guides
- Conjoint Analysis: Complete Guide -- Full methodology overview
- How to Design a Conjoint Study -- Design decisions that affect sample requirements
- 7 Common Conjoint Mistakes -- Including insufficient sample size
- MaxDiff Sample Size Requirements -- Sample needs for a simpler alternative
- Conjoint Sample Size Calculator -- Calculate your required sample
- Sample Size Determination -- General sample size principles for research
Plan your conjoint study -- try Quali-Fi free for 14 days.