"How many respondents do we need?" is usually the first budget question in a conjoint study. The honest answer is that it depends on what you want to do with the results, but the published rules of thumb narrow it down quickly. This guide walks through them, shows how tasks and respondents trade off, and explains how Ordeen sets a default.
Why there is no single number
Two kinds of error shrink as a study grows. Sampling error is about who you asked: a small sample may not represent the market. Measurement error is about how precisely each part-worth is pinned down, which depends on how many times each attribute level was seen and traded off. More respondents reduce both. More tasks per respondent reduce only the second. The right sample is the smallest one that keeps both small enough for the decision you are making.
Rules of thumb
Johnson's rule: the floor
Rich Johnson's rule of thumb for CBC main effects is that the number of respondents, tasks and alternatives should satisfy
$$\frac{n \times t \times a}{c} \geq 500$$
where $n$ is respondents, $t$ is tasks per respondent, $a$ is alternatives per task (not counting None) and $c$ is the largest number of levels on any one attribute. Rearranged, the minimum sample is
$$n \geq \frac{500\,c}{t \times a}$$
For two-way interactions, $c$ becomes the largest product of two attributes' level counts (Orme, 2019, ch. 7). Orme stresses that 500 "was intended to be a minimum threshold", and that 1,000 or more appearances per level is better.
300 for robust work
Johnson's rule controls measurement error only. For results you will base pricing or product decisions on, Orme recommends at least 300 respondents when you are not comparing subgroups, and notes that most commercial studies use between 150 and 1,200 (Orme, 2019, ch. 7). Sawtooth's own rule of thumb is the same: 300 per study (Halversen, 2020).
200 per group or segment
If you want to compare groups, such as regions, customer tiers or segments found in the data, each group needs about 200 respondents of its own. Four segments therefore means at least 800 (Orme, 2019, ch. 7; Halversen, 2020).
Exploratory work
For early-stage exploration where you only need the direction of preferences, Orme notes that 30–60 respondents can be enough. Treat those results as hypotheses, not prices.
More tasks or more people?
Within limits, they substitute for each other. Johnson and Orme (1996) found that at 300–400 respondents, doubling the number of tasks each person completes improves precision about as much as doubling the sample. The limit is respondent fatigue, and only more people reduce sampling error, so tasks cannot replace respondents when you need to represent a market or compare segments.
ACBC and small samples
Adaptive CBC collects more information per person, because the screening section and the tournament focus on concepts the respondent would actually consider. Orme (2019) notes that when the sample is particularly small and the attribute list is long, adaptive methods may be the better tool. See our guide to CBC vs ACBC.
MaxDiff sample sizes
For MaxDiff, the design question is how often each item is shown. Sawtooth recommends showing each item at least three, and ideally up to five, times per respondent when estimating with Hierarchical Bayes. With $K$ items shown $k$ at a time, that means at least $\lceil 3K / k \rceil$ sets per respondent (Sawtooth Software, Lighthouse Studio manual). The 300 overall and 200-per-group guidelines apply to the number of respondents.
Check the design before fieldwork
Sample size cannot rescue a weak design. Before launch, check that every level appears a similar number of times (level balance), that attributes vary independently (orthogonality), and that the design contains no impossible combinations. Ordeen scores each design on all three and shows the result before you launch.
How Ordeen sizes your study
Ordeen suggests 900 completes by default. That number is driven by segmentation rather than by the part-worths: it gives up to three segments at least 300 respondents each, so segment-level results are stable enough to act on. Ordeen also adds a buffer for respondents removed by quality checks. If you don't need segments, you can lower the target on the Field tab; the default is a recommendation, not a gate.
Worked examples
A simple CBC. Six attributes, the largest with 5 levels, 12 tasks of 3 alternatives each. Johnson's floor is 500 × 5 ÷ (12 × 3) ≈ 70 respondents. The 300 rule governs, so plan for 300 plus a cleaning buffer.
Comparing four customer tiers. Same design, but each tier needs about 200: plan for at least 800 plus buffer, with quotas so no tier falls short.
A price attribute with 8 levels. Johnson's floor rises to 500 × 8 ÷ 36 ≈ 112. The 300 minimum still governs, but if you also cut tasks to 8, the floor becomes 500 × 8 ÷ 24 ≈ 167. This is a reminder to keep the level count of your widest attribute in check.
MaxDiff with 24 items. Shown 4 at a time, each respondent needs at least 3 × 24 ÷ 4 = 18 sets. If that is too long, show 5 per set (15 sets) or split the list across respondents.
References
Orme, B. (2019). Sample Size Issues for Conjoint Analysis. Chapter 7 in Getting Started with Conjoint Analysis, 4th ed. Research Publishers LLC. Reprint (PDF)
Johnson, R. M. and Orme, B. K. (1996). How Many Questions Should You Ask in Choice-Based Conjoint Studies? Sawtooth Software Research Paper Series.
Halversen, C. (2020). Sample Size Rule of Thumb for Choice-Based Conjoint (CBC). Sawtooth Software blog. sawtoothsoftware.com
Sawtooth Software. Designing a MaxDiff Study. Lighthouse Studio Manual. sawtoothsoftware.com
