In the analysis of digital experiments, causal effect heterogeneity is often summarized into a few subgroups, as if that number were a well-defined estimator. However, beyond populations with latent classes, the 'true' number of subgroups depends on the model, not the population. To overcome this limitation, we propose the resolution profile, a new population functional that, for each fraction of explained heterogeneity, returns the minimum number of groups needed. This estimator is valid for any population, even without latent structure, and its inference is performed via a moment process with a Bayesian bootstrap adjusted by influence functions.
In the business context, applying this approach allows companies to understand how different customer segments respond to a campaign or product change without forcing an arbitrary number of groups. Q2BSTUDIO, as a software and technology development company, integrates this methodology into AI and custom software solutions, enabling its clients to uncover real causal patterns. The cloud infrastructure on AWS/Azure supports massive processing of experiment data, while cybersecurity ensures the protection of sensitive information. AI agents can automate the search for optimal resolution thresholds, and Power BI dashboards visualize heterogeneity profiles interactively.
The uncertainty in the number of subgroups is not a model selection problem but a threshold non-regularity: the resolution profile is an integer value that depends on a continuous path and presents discontinuities at certain points. At those knots, no pointwise selector is locally consistent, but a report based on simultaneous confidence bands maintains uniform validity. For businesses, this means they can rely on decisions about how many segments to consider without fear of statistical instability. Q2BSTUDIO offers consulting and development services that implement these methods, combining statistical rigor with modern cloud, AI, and business intelligence tools, so teams make decisions based on robust causal evidence.
