Hierarchical Bayes estimation
Hierarchical Bayes: utilities for every single respondent
Hierarchical Bayes (HB) estimates more than an average in conjoint and MaxDiff studies: it estimates utilities for every respondent. To do so, it borrows information from the whole sample and shows how differently your target group really decides – with an honest measure of uncertainty.
Building and testing is free. To publish, you choose a paid plan.
How Hierarchical Bayes works
What is Hierarchical Bayes?
Hierarchical Bayes is an estimation method for choice experiments such as conjoint and MaxDiff. Each respondent answers only a handful of tasks – too few to pin down their preferences reliably on their own. HB solves this by estimating separate utilities for each person while anchoring them to the distribution of everyone else. The result is individual part-worths per level or per item, which you can turn into importances, segments and market simulations.
Two levels: population and individual
The model has two levels. The upper level describes how utilities are distributed in the population – a mean and a spread for each parameter. The lower level holds each individual, whose utilities are treated as a draw from that distribution. People who answered little or inconsistently are pulled more strongly towards the centre; people who chose clearly and consistently keep a profile of their own. This “borrowing strength” is what makes stable individual estimates from only a few tasks possible at all.
Why not simply use a logit model?
An aggregate multinomial logit estimates one single set of utilities for everyone. With a heterogeneous target group, that average hides opposing camps: if one half barely notices the price and the other is highly price-sensitive, the logit shows a medium price sensitivity that fits nobody. On top of that comes the IIA property: a new offer takes share from all existing ones in proportion, even if it closely resembles just one of them. When you simulate with individual utilities, similar offers mainly compete with each other – and the market shares become more realistic.
Draws instead of a single number
HB is estimated with Markov chain Monte Carlo (MCMC). Instead of a point estimate, you get many draws from the posterior distribution: plausible values for every parameter, given the data and the model. From them you can read means as well as ranges of uncertainty – for a single utility just as for a simulated market share. The more the draws spread, the more carefully you should interpret a result.
Checking convergence
An MCMC run can only be trusted if the chains have found the same distribution. Three diagnostics are common: R-hat compares the spread within and between chains and should be close to 1. The effective sample size (ESS) tells you how many independent draws the correlated draws are worth. Divergences flag places where the sampler could not explore the distribution cleanly. If any of these values looks off, the result should not go into a report unchecked.
Hierarchical Bayes in 121surveys
For conjoint and MaxDiff
You start the HB run right in the results of a conjoint or MaxDiff question. It estimates utilities per respondent and derives the results you already know from the question – part-worths and importances for conjoint, the ranking of items for MaxDiff.
Runs in the background
HB never runs while you wait for a page; it runs in a separate background process. Depending on the size of the study, a run takes minutes to hours, and a progress display names the current phase – loading data, drawing, saving. You can keep working in the meantime and come back later, and a second click doesn't start a duplicate run.
With your current filters
The run uses exactly the responses your results currently show: filters, date range and the exclusion of test submissions apply just as they do in the view. That way you can estimate HB for a single target group or a single fieldwork period, for example.
Diagnostics at a glance
For every run, the results show R-hat, the effective sample size and the number of divergences. If a run doesn't meet all diagnostics, 121surveys marks the result as not reliable instead of quietly presenting it like a good one.
Draws for portfolio optimisation
For conjoint, an HB run stores 100 thinned draws per respondent. Portfolio optimisation builds on them: it looks for up to five products with the largest reach or the highest contribution margin against your base market and the no-buy option, and rates the best proposals with a 95% interval. The draws are deleted after 90 days.
Quick estimates first, then HB
For conjoint, the results start with counts and switch to a logit model once there are enough responses; above a minimum number, they point you to HB. The method advisor in the editor and in the results tells you both thresholds. For MaxDiff, you first see the count-based results and start HB on top when you need individual values.
Frequently asked questions
How many responses do I need for Hierarchical Bayes?
It depends on the design: the number of tasks per respondent, attributes and levels. The more parameters the model has to determine for each person, the more information it needs. 121surveys shows the threshold from which HB is offered in the method advisor and calculates a minimum sample size from your design for conjoint and MaxDiff. Fixed rules of thumb that ignore the design say little.
Are HB results always better than logit?
Not automatically. With a homogeneous target group, both are close and the logit is faster. HB shows its strength when people differ clearly and you want to build segments or simulate market shares. An HB run with poor convergence, on the other hand, is worse than a clean logit – which is why the diagnostics belong with every result.
Why does an HB run take so long?
MCMC draws many thousands of times from the joint distribution of all parameters – with every draw, a whole set of utilities for every respondent. In large studies with many attributes, that adds up to a great many values. 121surveys computes this in a separate process so the rest of the application stays fast, and shows you the progress.
What does “not reliable” mean in the results?
At least one convergence diagnostic – R-hat, effective sample size or divergences – is outside its limit. The figures are still shown, but marked. Then check whether there are enough responses, whether the design has too many parameters for the sample or whether a filter makes the data base too small, and run it again.
Can I run HB for a subgroup only?
Yes. The run takes over the filters of the results: set filters or a date range first, then start HB. Keep in mind that a small subgroup also gives the upper level of the model less information – the diagnostics show you whether it's enough.
Individual preferences instead of an average
Create a conjoint or MaxDiff question, collect responses and start HB right in the results – with a progress display, diagnostics and, for conjoint, portfolio optimisation based on the draws.
