Every behavior change program faces the same uncomfortable arithmetic. Budgets are finite, staff time is limited, and the number of people who could theoretically benefit almost always exceeds what any intervention can serve. The question is not whether to make choices, but how to make them defensibly.

For decades, the default answer was to spread resources evenly or to prioritize those with the greatest apparent need. Both approaches feel fair, but neither maximizes behavioral impact. Experimental evidence increasingly shows that the people who need help most are not always the people who respond most to a given intervention.

This is the terrain of behavioral targeting: using experimental methods to identify who benefits from what, and then allocating resources accordingly. It is a technically demanding practice with genuine ethical weight. Done well, it multiplies the impact of scarce resources. Done poorly, it can quietly widen the gaps it was meant to close.

When a randomized trial reports that an intervention increased physical activity by twelve percent, that number describes the average participant. But averages routinely conceal enormous variation. Some participants may have doubled their activity. Others may have shown no change, or even decreased. The single headline number is the mean of a distribution that can span from harm to transformation.

This variation is called heterogeneous treatment effects, and it is the rule rather than the exception in behavioral research. A smoking cessation program might work brilliantly for socially embedded smokers but fail for those who smoke to manage anxiety. A financial literacy nudge might shift behavior among the moderately organized but bounce off both the highly disciplined and the deeply overwhelmed.

Detecting this variation requires more than post-hoc subgroup analysis, which is notoriously prone to false positives. Modern approaches use pre-registered moderator analyses, causal forest methods, and factorial designs that systematically vary intervention components. These techniques estimate not just whether something works, but for whom and under what conditions.

The practical upshot is that replication of an average effect tells you less than you think. Two studies can report identical mean effects while the underlying response distributions differ completely. Intervention designers who ignore this variation end up delivering programs that succeed on paper while failing the people they were meant to reach.

Takeaway

The average treatment effect is a starting point, not a conclusion. The real question is not whether an intervention works, but for whom, how much, and why the response distribution takes the shape it does.

Once you accept that responses vary, the next question is how to identify likely responders before you spend resources on them. Experimental approaches to targeting fall into several families, each with distinct trade-offs.

The simplest is risk-based targeting: prioritize those with the greatest baseline need. This is intuitive and often defensible, but it conflates need with responsiveness. A patient with severe hypertension may benefit less from a medication adherence text message than a patient with moderate hypertension, simply because their case requires more intensive support.

A more sophisticated approach is uplift modeling, which uses experimental data to predict each individual's expected treatment effect based on their observable characteristics. Rather than asking who is worst off, it asks who is most likely to change because of this intervention. Studies in domains from tax compliance to vaccination outreach have shown that uplift-targeted campaigns can produce two to five times the behavioral change of untargeted or need-targeted approaches at equivalent cost.

A third option is adaptive assignment, where interventions are tested and refined during rollout. Multi-armed bandit designs and sequential multiple assignment randomized trials allow programs to learn from early participants and route later participants toward the components that work best for people like them. This turns implementation itself into an ongoing experiment.

Takeaway

Targeting is not about picking the neediest or the easiest. It is about matching intervention components to the people whose behavior is most likely to shift because of them, using evidence rather than intuition.

Targeting the most responsive individuals is efficient, but efficiency is not the only value at stake. If a smoking cessation program consistently produces larger effects among higher-income participants, uplift-based targeting will direct resources upward, potentially widening health disparities even as it maximizes aggregate quit rates.

This tension between efficiency and equity has no clean resolution, but it can be managed transparently. One approach is constrained optimization: maximize expected behavioral impact subject to distributional constraints, such as requiring proportional representation across demographic groups or minimum service levels for high-need populations.

Another is to examine why response varies. If some groups respond less to an intervention, the honest question is whether the intervention itself was designed for them. Low responsiveness often reflects poor fit, not fixed individual traits. Investing in intervention variants tailored to underserved groups may produce more equitable and larger overall effects than targeting the existing version more aggressively.

Transparency matters as much as method. Participants and stakeholders deserve to know how targeting decisions are made and on what data. When targeting logic is hidden, it can drift into patterns that would not survive scrutiny. When it is documented and reviewed, it becomes a tool for accountability rather than a source of quiet bias.

Takeaway

Targeting is a value-laden design choice, not a neutral technical operation. The algorithm reflects the priorities of whoever built it, and those priorities deserve to be stated openly.

Behavioral targeting is neither a technical shortcut nor a moral hazard. It is a discipline that forces designers to confront what they actually know about how their interventions work and for whom.

The practical recommendation for intervention designers is to plan for heterogeneity from the start. Pre-register moderator analyses, collect the covariates needed to estimate individual treatment effects, and build adaptive elements into rollout when possible.

The broader recommendation is to treat targeting decisions as part of the intervention itself, subject to the same scrutiny as any other design choice. When targeting is done with rigor and transparency, limited resources reach further and the people served are better matched to what they receive.