· The Fieldbase team choice-modelingcbcmaxdiffmarket-research

CBC vs MaxDiff — when to use which

Choice-based conjoint and MaxDiff look similar on the surface, but they answer different questions. Here's the plain-English difference, the three cases where each one wins, and a decision rule you can use on your next brief.

A client asks: “we want to know what matters most to our customers.” You could run choice-based conjoint. You could run MaxDiff. The two methods look similar — both ask respondents to make trade-off choices, both produce a ranked list of importances — but they answer different questions. Pick the wrong one and the output doesn’t fit the brief.

Here’s the plain-English difference, and a decision rule you can apply on your next project.

The core distinction, in one sentence

CBC measures the value of features that bundle together into a product. MaxDiff measures the relative importance of items in a flat list.

That’s the whole thing. The rest is detail.

If what you’re testing can be described as a product with attributes — a laptop with price, screen size, weight, battery life; a bank account with rate, fees, perks; a car with brand, engine, color — CBC fits. The attributes trade off against each other inside a single purchase.

If what you’re testing is a flat list of things where nothing bundles — reasons someone buys a brand, benefits of a product, messages for an ad campaign, features for a roadmap, concerns about a service — MaxDiff fits. The items compete for mindshare, not for bundle slots.

When CBC wins

1. You need to simulate a market

If the client wants to answer “what share of preference does our product get versus competitors at different price points?”, you need utilities for every attribute level. CBC gives you a utility per level (e.g. a utility for $999, $1,199, $1,399) that you can plug into a share-of-preference simulator. Change a level, watch share move.

MaxDiff can’t do this. It gives you importance scores for items; it does not give you a utility for a specific price point. If the deliverable is a simulator, run CBC.

2. You care about price elasticity

CBC naturally handles price as an attribute with multiple levels. You can derive willingness-to-pay directly from the price utility curve.

MaxDiff can rank “lower price” against other benefits, but it won’t tell you how much lower.

3. Interactions matter

If brand effects depend on price (your brand commands a premium only up to a certain point), or if feature preference depends on brand (one feature matters for premium buyers but not budget buyers), CBC can model those interactions with the right design and estimation. MaxDiff doesn’t model interactions at all — it assumes each item’s importance is independent of the others.

When MaxDiff wins

1. You have 15+ items and you want a clean ranking

CBC breaks down fast when attribute lists get long. A single-attribute CBC with 20 levels requires either a huge design or forces respondents through many tasks. MaxDiff was designed for exactly this — give it 20 items, show subsets of 4 or 5 per task, and respondents compare them one task at a time.

Use cases where this fits:

  • Ranking 20 potential product features for a roadmap
  • Ranking 15 ad concepts
  • Ranking 12 reasons-to-believe for a brand
  • Prioritizing a list of policy concerns

2. The items don’t bundle

Nobody buys a “feature” or an “ad concept” on its own — these are flat lists where only one thing (or a few things) wins budget attention. CBC’s model (what bundle do you prefer?) doesn’t fit. MaxDiff’s model (which of these matters most?) does.

3. You want a direct, defensible score

MaxDiff scores are easy to communicate. “Item A has an importance score of 15.2, Item B is 12.8.” They can be rescaled to sum to 100 or converted to preference share. The chart is a bar chart with a clear winner. Executives like them.

CBC utilities are harder to explain to a non-technical stakeholder. “The part-worth for Brand X at the $999 level is 1.4” means nothing without the simulator sitting next to it.

The hybrid pattern — MaxDiff pre-screen → CBC

Some projects use both. A pattern worth knowing:

  1. Start with MaxDiff over 20 potential features/attributes to identify the top 5-6.
  2. Design a CBC with those top attributes as the design matrix.
  3. Run both in the same survey (MaxDiff first, CBC second with piping).

The MaxDiff tells the client what to put in the CBC. The CBC then quantifies the trade-offs. You get both a feature-prioritization story and a market simulator from one survey.

This takes more design effort — longer survey, careful piping, bigger sample for the CBC to be statistically precise after the pre-screen — but when the client is early-stage (“we have ten ideas and don’t know which to put in a product”), it’s the right flow.

The decision rule

Three questions, in order:

  1. Does the thing being tested have attributes that trade off within a single purchase or choice? If yes → CBC. If no → MaxDiff.
  2. Does the client need a simulator or price-elasticity output? If yes → CBC (even if the item count is high; use an alternative-specific CBC or partial-profile design).
  3. Is the list of things 10+ items long, with no attribute structure? If yes → MaxDiff.

If you’re unsure, write out what the client’s deck slide looks like at the end of the project. If it’s a bar chart ranking items → MaxDiff. If it’s a share-simulator or elasticity curve → CBC. If both → hybrid.

Common mistakes

Running CBC on a list of flat items because “it feels more sophisticated.” CBC on items that don’t have attribute structure wastes respondent capacity on fake trade-offs. The utilities you get out are noisy and you can’t simulate anything.

Running MaxDiff when you actually need pricing. MaxDiff on a list that includes “lower price” or “better value” doesn’t give you a price curve. You get an importance score for the concept of price, which the client can’t act on.

Underpowering MaxDiff. MaxDiff needs tasks per respondent to produce stable scores — typically each item should appear 3-4 times across the respondent’s tasks. For 20 items with 4 per task, that’s ~15 tasks per respondent. Shorter is unreliable.

Overpowering CBC with too many attributes. A CBC with 7 attributes at 5 levels each is theoretically fine but asks a lot of respondents. Keep attribute count to 5-6 for best data quality, or use partial-profile designs.

How Fieldbase handles both

CBC and MaxDiff are first-class in Fieldbase. The Design Test runs before fielding — Design Health (level exposure, prohibition flags, version coverage) and Precision (per-level standard errors, D-efficiency, implied-N) scorecards sit next to the design in the editor so you know before you buy sample whether the design will produce usable estimates. Aggregate MNL part-worths and MaxDiff best-worst scores run in-app; the raw cbc_long and maxdiff_long datasets export cleanly for Hierarchical Bayes in Sawtooth, R, or Python.

See the choice modeling page for the design-asset workflow and the pre-fielding Precision check.

Fieldbase is AI-native survey software for research programmers. CBC, MaxDiff, and Design Test are built in; quotas hold under concurrent traffic; exports drop straight into SPSS, Stata, and Excel. Currently in early access — get in touch.