The Goldilocks Zone: Matching Goals to Interventions

The Goldilocks Zone: Matching Goals to Interventions

Why the size of your goal must match the size of your intervention, what nudge meta-analyses really tell us about effect sizes, and how to balance exploration and exploitation when resources are scarce.

The brief

A regional health provider wants more people to attend their screening appointments. Attendance is 60 percent. The board has announced a target of 85. A behavioural team has been given a quarter, one analyst, and a budget that covers text messages and not much else. Everyone in the room believes the target is reachable, and nobody has checked. This paper is about the check, and it will keep coming back to this programme, because almost everything that goes wrong in behaviour change goes wrong in the gap between the number on the slide and the resources under it.

Goals and barriers are the same object seen from two sides

Between the screening population and the outcome sits a stack of barriers. The team maps six that matter: patients do not receive the letter, do not read it, forget the date, cannot get time off work, cannot get transport, and are anxious about the result. Each layer removes some fraction of the people who would otherwise attend. The stack is why attendance is 60 and not 100.

A goal is a statement about how many people must get through the stack. Set the goal at 64 percent and the team may only need to remove the one barrier that is stopping the most persuadable group, the people who simply forgot. Set it at 85 and the team must remove nearly every layer, because any one left standing stops enough people to keep the number short.

This sounds obvious written down. It is violated constantly, because goals are set by one group (boards, funders, strategy) and interventions are designed by another (delivery teams, agencies, behavioural units), and the two rarely sit down to check whether the barrier stack the second group can attack is large enough for the target the first group has announced. In the screening programme, the 85 was set before anyone had drawn the stack.

What a text message is actually worth

To check the match, the team needs honest numbers about what interventions deliver. The nudge literature provides the cleanest case because it has been meta-analysed several times with conflicting results, and the conflict itself is instructive.

The DellaVigna and Linos result is the most useful because it is a complete record: every trial two US government nudge units ran, published or not. On a baseline take-up of roughly 17 percent, the typical nudge added 1.4 points. Publication bias and low statistical power explained about 70 percent of the gap between that figure and what appears in journals. The remainder came from differences in the interventions themselves: academic studies more often tested defaults and in-person interventions, while the units mostly tested letters, emails and reminders. Letters, emails and reminders are exactly what the screening team can afford.

The contrast with defaults is the important calibration: Jachimowicz and colleagues' pooled default effect of d = 0.68 is an order of magnitude above the bias-corrected nudge estimates. Benartzi and colleagues (2017) make the complementary point that even small nudge effects can be highly cost-effective per unit spent, which is an argument for using them at scale, not for expecting them to deliver transformation. The practical reading is not "nudges do not work". It is that a typical, cheap, communication-style nudge, deployed at scale, moves a behaviour by a small single-digit number of percentage points, and that anyone promising more from that class of intervention is quoting a biased literature. Around two percentage points is a sensible planning figure. That is a real, cost-effective gain when the base is millions of patients. It is nowhere near a transformation goal.

The two ways to miss the zone

Too hot: a large goal on a small intervention

This is the screening programme as briefed, and it is the common failure. The team designs a well-crafted reminder, runs it, and reports a statistically significant 2.6-point improvement to 62.6 percent. Everyone is disappointed. The intervention was fine. The match was wrong. Worse, the disappointment is usually charged to behavioural science ("we tried nudging, it did not work") rather than to the mismatch, and the organisation loses a tool that would have been perfect for a goal of the right size.

Too cold: a small goal on a large intervention

The rarer but costlier failure, and the screening programme's neighbour down the corridor has it. A modest outcome, getting a few more people to return a consent form, is attacked with a redesigned process, a training programme, new incentives and a change-management workstream. The barrier stack for that outcome had one meaningful layer. Removing it with a default or a reminder would have done the job at a hundredth of the cost. The over-built intervention succeeds and is celebrated, and nobody notices the waste because success hides it.

Turn the logic around: let resources set the goal

Goal-setting usually runs from ambition to plan. The Goldilocks logic also runs in reverse, and the reverse direction is often more honest. If the resources available for a solution are small, the goal for its impact should be small. A team with one analyst, one text-message channel and no authority over the appointment system should promise a two-point gain and deliver it, rather than promise 85 and deliver 62.6.

This is not defeatism. Small, reliable, cheap gains compound. A unit that delivers a two-point improvement on twenty behaviours a year is worth more than one that promises a twenty-five-point improvement on one behaviour and misses. And a track record of matched promises is what earns the authority and budget needed to attempt the structural changes that deliver large effects. The path to big goals runs through correctly sized small ones.

Exploration and exploitation: allocating what you have

Suppose the screening team wins a larger budget on the strength of its honest two points. The next question is how to spend it across approaches, and here decision theory offers a precise vocabulary. In a multi-armed bandit problem, a player faces several options with unknown payoffs and must decide, at each turn, whether to exploit the option that has performed best so far or explore a less-tested one that might be better. Exploit too early and you lock into a mediocre option. Explore too long and you spend the budget learning instead of earning.

Behavioural programmes face the same trade-off. Exploitation is deploying the intervention with the best evidence, here the reminder, and accepting its known, modest effect. Exploration is testing something new, such as a default appointment slot, an employer letter, or a mobile screening unit that removes the travel behaviour entirely, with a higher variance of outcomes and a chance of a much larger effect. Three principles from the bandit literature translate directly:

• Explore in proportion to the time horizon. A programme with one shot at a result should exploit. A programme that will run for years, as screening does, should explore heavily early, because a better arm found in year one pays off in every later year. Most organisations do the opposite: they explore in pilots that have no follow-on and exploit in long programmes that never test alternatives.
• Explore in proportion to the gap. If the best known intervention delivers two points and the goal needs twenty-five, exploitation cannot reach the goal, and the rational choice is to spend the budget looking for a higher-payoff arm. If the best known intervention already delivers the goal, exploration is a luxury.
• Explore cheaply and often, not expensively and once. Optimal bandit strategies test many arms with small samples before committing. In practice this means many small experiments at identified leverage points, run in parallel across a few clinics, rather than one large pilot of the option the team already preferred.

Risk appetite is what sets the balance. Exploration is risk-taking; exploitation is risk-avoidance. Neither is virtuous in itself. A programme that only exploits will never find the structural change that makes 85 reachable. A programme that only explores will never bank a result. The Goldilocks zone for resource allocation is a portfolio: enough exploitation to guarantee the modest goal, and enough exploration to have a credible chance at the ambitious one.

The brief, rewritten

Here is what the screening team takes back to the board. First, the goal restated in tiers: 64 percent is achievable this quarter with reminders, and that is the exploit arm; 75 percent requires a default appointment slot with easy rescheduling and an employer letter, and that is moderate exploration with a structural element; 85 percent requires transport partnerships and a redesigned results process, and is a two-year goal. Second, the budget allocated accordingly: most of it to the proven reminder, a meaningful share to the two structural experiments, a small share to the genuinely novel mobile-unit arm. Third, an agreement that the number the board hears this quarter will be 64, and that it is a success.

Placing this inside CLEAR

The Clarify step of the CLEAR framework exists to set objectives and key results before any intervention is designed. The argument here sharpens what that step must produce: not just a goal, but a goal that has been checked against the barrier map (Leverage) and the planning effect sizes of the interventions the team can actually afford. The Experiment step is the exploration budget. The Analyse and Refine steps are where the exploit-versus-explore balance gets updated with real data, exactly as a bandit algorithm updates its estimates after each pull. The companion paper on behaviour-free design describes the top of the intervention ladder, where the large effects in the table above come from.

Conclusion

Match the size of the goal to the size of the barrier stack you are willing to remove. Use honest planning numbers, about two points for a light-touch nudge, an order of magnitude more for defaults and structural change. Let small resources set small goals and be proud of meeting them. And when the gap between what you can prove and what you need is large, spend deliberately on exploration, because no amount of exploiting a two-point intervention will ever reach a twenty-five-point goal.

© 2026 Erik Bohjort · clear-framework.com