Uplift versus propensity: find accounts contact can help

Two questions can lead to different queues

A propensity model asks which accounts are likely to reach an outcome. A treatment-effect model asks whose outcome might change because of an intervention. An account already committed to buying can have high propensity and little incremental response to another campaign.

Define which question the programme needs answered. Capacity planning may need outcome prediction. Choosing where to add an expensive intervention may need evidence of incremental response. Do not label a conversion model as uplift because the term sounds more advanced.

What the methods paper contributes

Künzel and colleagues’ 2019 PNAS paper develops metalearners for estimating conditional average treatment effects. It explains approaches that use supervised-learning methods to estimate how effects vary across contexts.

The method does not create causal evidence from arbitrary CRM history. Identification assumptions, treatment assignment, overlap and adequate observations still matter. Our ABM application is to establish a credible comparison before using account features to estimate where an intervention is more useful.

Explore the original methods and findings in Metalearners for estimating heterogeneous treatment effects using machine learning.

The practical workflow

1. Choose prediction or causal effect
2. Specify treatment and baseline
3. Preserve assignment and features
4. Inspect overlap and subgroup size
5. Validate before allocating effort

Compare the approaches

Define the treatment and baseline

Specify exactly what is added: a bespoke workshop, an additional sequence or a coordinated content play. Keep the baseline programme explicit. If the intervention combines several changes, the experiment estimates the combined change rather than the separate contribution of each component.

Preserve account-level assignment when contacts can influence the same buying group. Record other seller activity and account maturity. Use features available before treatment. An opportunity stage updated after the intervention should not become an input pretending to predict the initial effect.

Read an illustrative subgroup carefully

Suppose a fictional randomised pilot suggests a larger meeting-rate difference among accounts with a recent public expansion. Before treating that as a targeting rule, inspect subgroup size, whether the analysis was planned and whether the pattern repeats in fresh data.

A small subgroup with a large apparent lift can result from chance. A business narrative explaining the result does not remove that uncertainty. Keep the finding as a hypothesis until the evaluation supports using it. Do not repeatedly search segments until one looks impressive and report only that segment.

Preserve the counterfactual evidence

Ghost-ad research illustrates how careful exposure and control design can improve advertising evaluation. It is a platform-specific method, but the broader lesson for ABM is to preserve assignment and eligibility before interpreting engagement.

If the only evidence is that engaged accounts converted more often, the result may describe selection rather than incremental effect. An uplift model built on that history can learn who was already active. Check the causal design with a qualified analyst before using the model to allocate substantial programme effort.

Explore the original methods and findings in Ghost Ads: Improving the Economics of Measuring Online Ad Effectiveness.

Choose a simpler evaluation when needed

For a small enterprise account universe, a clear experiment and an interpretable subgroup report may be more useful than a complex treatment-effect model. Start with the decision the sample can support. Keep operational learning separate from uncertain commercial estimates.

Use propensity for a named forecasting or prioritisation purpose. Use treatment-effect analysis when a defensible comparison and sufficient data support the question. Review cost per incremental accepted outcome before expanding an expensive intervention.

For the next part of this decision, read Advertising holdouts and ABM: preserve the counterfactual.

Your next-action checklist

• Propensity: Label the predicted event. Check the limitation: does not estimate contact effect.
• Average treatment effect: Start with a credible comparison. Check the limitation: may hide useful variation.
• Conditional effect: Validate in fresh accounts. Check the limitation: needs more data and assumptions.
• Engagement correlation: Avoid causal wording. Check the limitation: selection can explain it.

Use the comparison to choose a bounded next step. Record the evidence, the responsible owner, and the review decision before extending the play to additional accounts.

Questions this guide answers