Ordeen Choice Modelling Collapses several months into days— Without Sacrificing Statistical Rigour
The Problem with Traditional Conjoint
You have a question: "How much would my customers pay for 5G, and which brands would they walk away from at £100 over Apple?"
In the traditional world, finding the answer takes eight weeks. You brief an internal insights manager who briefs a research consultancy, who briefs Sawtooth (the licensed software), who briefs a panel partner. Quotes ping back. The brief is rewritten three times because the consultancy wants 250 respondents, the panel can field 400 cheaper, and the head of pricing wants a screener for "current iPhone owners only." A draft survey arrives in week four. Fielding starts in week five. Cleaned data lands in week six. The deck — built in PowerPoint, with charts re-drawn by hand from a Sawtooth export — arrives in week eight.
By then, the question has changed. The £100k+ spend has been paid. The findings sit in a slide that someone might open again next quarter.
The traditional process isn't broken because the people are bad. It's broken because every handoff is a meeting, every deliverable is bespoke, and the software that runs the actual analysis is priced like a luxury asset rather than a tool.
Our Approach: Collapse the Workflow, Keep the Mathematics
Ordeen Choice Modelling is the same brain — Hierarchical Bayes, Latent Class, MaxDiff, Adaptive CBC, MBC — wrapped in a workflow you can run yourself. The eight-week process collapses into three or four hours of guided clicks, because the things that used to require meetings (writing the survey, choosing the methodology, finding competitors for the simulator, naming the segments) now happen in the product. The mathematics stays exactly as rigorous as the textbook says it should be.
We use AI for the setup and storytelling parts of the workflow — proposing attributes, drafting screeners, suggesting competitive Market sets, naming segments after they've been clustered. We use textbook statistics for everything that touches the actual estimation: Stan-based Hierarchical Bayes for the part-worth utilities, Latent Class for the segmentation alternative, k-fold cross-validation for hit-rate honesty, Krinsky-Robb posterior draws for the 95% credible intervals you'd want on any number that touches a pricing decision.
The AI / Statistics Partnership
The split is deliberate.
AI excels at:
Drafting attribute suggestions from a one-line business question
Identifying realistic competitors covering ~80% of market share for the Simulator
Translating "Battery 99% importance, over-indexes on 30-hour" into the segment name "Battery-first power users"
Generating sentiment themes from open-ended responses
Statistics excels at:
Estimating utilities from forced-choice answers (Hierarchical Bayes — Sawtooth's textbook method, unchanged)
Quantifying uncertainty (95% credible intervals on every utility, share, and willingness-to-pay)
Detecting whether a finding is signal or noise (every level shown with t-statistics; non-significant cells go grey)
Validating the model out-of-sample (k-fold CV hit rate, not just in-sample fit)
AI proposes. Statistics disposes. We never let the language model touch the part-worth estimation, and we never let the cluster algorithm name itself without checking the name against the data it was clustered from. Every segment label is post-validated: if Gemini calls a cluster "Camera Lovers" but Camera importance is 0% for that cluster, the label is rejected and a deterministic name is substituted from the actual top driver.
A Concrete Example
A UK SaaS team needs to price ten add-ons (SSO, CRM Pro, Audit log, Mobile app, Premium support, Dedicated CSM, …). The traditional path: a £40k MBC study from a research vendor, eight weeks, twenty stakeholders.
The Ordeen path:
Brief (10 minutes) — paste the business question, accept the AI-suggested screener, confirm the country.
Design (30 minutes) — review the AI-suggested items, set prices, watch the live mobile preview.
Field (one click) — launch into our integrated Prodege panel. Real respondents start arriving in minutes. The only wait time is waiting for real respondents to complete the survey in days — this is where the insight is derived so well worth the wait.
Results (automatic, when fielding completes) — segmented by k-means and Latent Class; per-segment importance and utility tables with significance colouring; Price-Value Score per item, marked green for under-priced and red for over-priced.
Simulate (interactive) — drag prices up and down, see share-of-preference recompute live, run elasticity sweeps per segment.
Handling The Hard Cases
Small samples — Hierarchical Bayes shares information across respondents, so a 300-respondent study still produces stable individual-level part-worths. The traditional "we need 1,000 to be safe" rule was a relic of when MCMC was a research artifact, not a product.
Many attributes — when usable attribute count exceeds 6, the system auto-routes from CBC to Adaptive CBC (ACBC), running the BYO → screener → tournament three-phase flow that Sawtooth charges a separate licence for.
Bundle pricing (MBC) — Menu-Based Choice for "which add-ons would you buy at these prices?" decisions, with item-level Price-Value Scores that surface mispriced items at a glance. The MBC editor refuses to silently estimate a study with a 50× price spread (a single MBC price coefficient mathematically can't fit that) — it warns the seller up-front, the way a research consultant would.
Mixed-method comparison — every project ships with both k-means (fast, descriptive) and Latent Class (slower, textbook-statistical) segmentation. Sellers compare. The chip strip on the Results tab lets them filter every section by either algorithm's segment.
What This Means for Your Business
Days, not months. Brief through Results in a days for studies that used to take multiple months.
Cost that scales with usage, not vendor day-rates. Pay for panel respondents and a monthly platform fee. No statements of work, no scoping calls, no per-deck billing.
Methodology that survives scrutiny. The estimation stack is Stan-based Hierarchical Bayes and Latent Class — the same methods cited in the original Orme & Lyon papers. We didn't reinvent the maths; we just removed the friction wrapped around it.
Findings that share themselves. A public read-only Results link, scoped to the segment you care about. Email it to a stakeholder. They see exactly what you see.
One Full Report CSV. One row per respondent, every demographic, every utility, every z-score, every k-means and Latent Class assignment — drops directly into R, Python, Stata, or an internal data warehouse.
The Bottom Line
We've streamlined the process. The statistics underneath are unchanged — we use the same Hierarchical Bayes the textbook describes, the same Latent Class the literature validates, the same significance tests a quant would expect. What we changed is everything around the statistics: the brief, the design, the fielding, the segmentation naming, the share-of-preference simulator, the sharing of results.
