Profile segmentation and adoption modelling

Segmenting with a sound method identifies subgroups with homogeneous behaviour, risk profile or expected response within a population that hides that heterogeneity when it is treated as a single block. This applies when you have to decide how to allocate an intervention across a population, such as patients, beneficiaries or areas, that does not respond in the same way throughout. Spreading effort evenly across a population that is not homogeneous can misallocate resources: some subgroups adopt or respond early, others late and others not at all. A data-derived segmentation makes those boundaries explicit with a reproducible criterion, instead of setting them by eye or by inherited administrative categories. Knowing the population qualitatively is not the same as having segments that are data-derived, validated and actionable. What the service provides is the step from an intuition about heterogeneity to a segmentation you can use to focus the intervention.

Segment boundaries are not set by eye: they are derived with unsupervised methods and their stability is assessed, so that they do not depend on a single criterion or on a one-off partition of the sample. This applies to all segmentation, where the risk is producing groups that look clear but do not hold up. The algorithm is chosen to suit the structure of the data, for example partitioning, hierarchical or latent class methods, and the segmentation is built by triangulating a data-derived method with an interpretable one, so that the classification is both statistically consistent and defensible to whoever makes the decision. The stability of the segmentation is then assessed, instead of accepting a classification that only holds on the sample that produced it. This is precisely how subjectivity is reduced: objective criteria, two methods tested against each other and a check that the segments are not an artefact of the data. What you receive is a reproducible, auditable segmentation, not a partition set by personal judgement.

Segmenting describes the present, that is, which subgroups exist and how they differ. The adoption model projects expected behaviour: which profiles are most likely to respond, in what order and to what degree as penetration increases. This applies when classifying the population is not enough and you need to anticipate its response to an intervention that has not yet been rolled out. The adoption model estimates the propensity to respond by profile, choosing the family that suits the phenomenon, for example models for excess zeros when an excess of non-response is expected or observed, and combines classification with probabilistic scoring to rank the population by propensity. Segmentation is the descriptive base; the adoption model is the predictive layer built on top of it. One honest limit applies: the model projects behaviour under calibrated assumptions; it does not predict the future with certainty. Describing who the population is and forecasting how it will respond are two steps, and the value lies in linking them.

By anchoring the model to external reference sources when they are available: a response model that is not calibrated against external sources is an elegant hypothesis, not a defensible projection. This applies to any adoption model that will be used to size resources or decide on a roll-out. The weights are calibrated against census records, administrative records or official statistics, and the projections are anchored to the penetration or response reference figures provided by the client, so that the curve is more consistent with the target population than with an idealised one. In addition, a saturation curve governs the response at each level of penetration, instead of assuming instant adoption, which is rarely defensible for a real intervention. Calibration against external anchors is what separates a credible projection from a well-presented optimistic assumption. What you receive is a model calibrated against the reality of the client's population, with its references made explicit, not a theoretical curve that assumes the best-case scenario.

Yes: building on the segmentation and the adoption model, the expected effect of the intervention can be estimated, parameterised by its configuration: which variant, what intensity and what coverage. This applies when the decision is not only how many people will respond, but what happens with each possible design of the intervention before committing to one. The effect is estimated by segment, not just as an aggregate average that hides heterogeneity, and the outputs are built to be comparable across alternative configurations. This makes it possible to contrast, for example, a more intensive variant reaching fewer people with a lighter, broader one, and to see the expected effect of each in each subgroup. A note on scope: this "parameterised effect" is the effect of the intervention according to its configuration, not a health economic cost-utility evaluation with QALYs or DALYs, which is a different service. What you receive is a comparison of configurations by their expected effect in each segment, so that you can decide on the design before rolling it out.

Profile segmentation and adoption modelling answer who responds and in what order: they identify the subgroups and project their propensity to adopt the intervention. Ex ante modelling of interventions estimates the potential aggregate effect of an intervention on an indicator before it is rolled out, integrating the available sources. They do not answer quite the same question. This service focuses on heterogeneity in the response, the map of who adopts, when and how much; ex ante modelling focuses on the expected change in the target indicator. They complement each other naturally: the segmentation and the adoption curve can be an input to ex ante modelling, providing the detail of how the response is distributed across the population. If the question is "which subgroups respond, and in what order?", this is the service; if it is "what total change can we expect in the indicator?", it is ex ante modelling. One maps the response; the other estimates the effect on the outcome.

It depends on the focus. If what matters is the response to an intervention and its adoption over time, profile segmentation and adoption modelling is the right fit; if the question is clinical and the aim is to group patients by risk to guide management or follow-up, the natural framework is patient stratification. The distinction matters because both share grouping techniques but answer different questions. Patient stratification focuses on classifying by clinical or prognostic risk to inform care decisions; segmentation and adoption modelling focus on who will respond to an intervention and in what order as its penetration increases. The two may overlap in a specific project, but the deciding question is whether the aim is clinical risk or the propensity to respond. If what you need is to stratify by risk for clinical practice, that is what patient stratification models are for; if it is to project the adoption of, or response to, an intervention, it is this service.

You need data on the population to be segmented, meaning the characteristics of the individuals, areas or units, and the external references that allow calibration: administrative records, official statistics or any penetration or response anchors the client has. The client rolls out and distributes the intervention and, where applicable, measures actual adoption. The service segments the population, selects the predictors with objective criteria (for example variable selection methods or regularisation), estimates the expected propensity to respond and calibrates the projections when there are enough external references. Building the variables that control for heterogeneity and preparing the predictors are part of the work before segmenting. ADI does not measure the adoption that occurs after roll-out, since the client records it; what ADI provides is the segmentation and the projected response before roll-out. What you receive is the segment map and the calibrated adoption model; the roll-out and the decision remain the client's.