Nudge

A nudge is any feature of how a choice is presented that changes behavior in a predictable direction without removing options or altering their costs. The attribution is split, and saying so is the point. Cass Sunstein and Richard Thaler made the argument in "Libertarian Paternalism Is Not an Oxymoron" in the University of Chicago Law Review in 2003, alongside a companion piece in the American Economic Review. That article uses the word "nudge" exactly once, as an ordinary verb, and does not contain the phrase "choice architecture" at all. Both became terms of art in the 2008 book Nudge, which is where the noun and the design vocabulary come from. What the 2003 article does supply is the standard a nudge has to meet: libertarian paternalism, in its authors' words, "is a relatively weak and nonintrusive type of paternalism, because choices are not blocked or fenced off," and in its most cautious forms it "imposes trivial costs on those who seek to depart from the planner's preferred option."

The two halves of the name each do work. The paternalist half is the claim that some presentations lead to better outcomes than others and that whoever designs the choice cannot avoid choosing one. The libertarian half is the constraint: the option a person would otherwise have taken must remain available at the same price and at negligible cost to reach. That constraint is what excludes most policy tools. A tax changes a price. A mandate removes an option. A financial incentive changes the arithmetic of the decision. None of them is a nudge, and calling them one drains the term of the only content it has.

In practice the family is small and recognizable: a default that applies when nobody acts, a simplified form, a reminder timed to a deadline, a prompt that asks someone to say when and where they will do something, a comparison with what similar people do, and the ordering or prominence of options on a page. The default is the strongest member of the family and has its own mechanics. Everything else in the list works by reducing the effort or the forgetting between an intention and the act.

The honest effect-size picture is this page's most useful content, because the gap between the published literature and practice is unusually well measured. Stefano DellaVigna and Elizabeth Linos assembled every trial run by two of the largest government behavioral units in the United States, 126 randomized trials covering more than 23 million people, and compared them with a matched set of nudge trials from academic journals. In the academic sample the average intervention raised take-up by 8.7 percentage points, a third above the control rate. In the government trials the average was 1.4 percentage points, about one sixth as large, and still clearly different from zero. Their conclusion about the gap is as important as the numbers: publication bias in the journals, made worse by small samples, can account for the whole difference. Two details of that work should travel with the figures. Interventions that changed a default were excluded from the main analysis, so these numbers describe the weaker members of the family rather than the strongest one. And the government trials were mostly letters and forms, which is a real constraint on what a public agency can do rather than a failure of imagination.

Reviewing the field in 2017, the Royal Swedish Academy of Sciences noted that nudges compare favorably with traditional policy tools when impact is measured against cost, while quoting the same researchers' caveat that more calculation is needed before the comparison is settled. Cheap and small is a perfectly respectable result. It is a different result from the one the popular literature reports.

The objections are worth stating in their own terms rather than as a formality. Robert Sugden and colleagues argue that the approach treats departures from the standard rational model as mistakes for a policy-maker to correct, which assumes the policy-maker can know which underlying preference is the real one. Ayala Arad and Ariel Rubinstein report evidence that some people object to being steered at all, independently of whether the steering leaves them better off, which is a cost the usual accounting does not capture. And there is a consumer-side point that follows directly from the definition: choice architecture is not a government technique. Every form, checkout and enrollment screen has an architecture, and the same instruments work exactly as well when the preferred option is the one that suits whoever built the page.

Designing one follows a fixed sequence. Identify the moment the choice is actually made. Work out what happens if the person does nothing, since that outcome is already being chosen for them. Change the presentation without touching the options or their prices. Then test it against a control group, because the field's own record shows that intuition about effect sizes is poor in both directions.

A hypothetical, built on the measured averages rather than on an invented one. An agency writes to 200,000 people it believes are eligible for a benefit, and 17 percent of a control group claims it, which is close to the average control take-up in the government trials. A nudge of the size those trials average, 1.4 percentage points, brings 2,800 additional claims (200,000 × 0.014). Judged instead against the 8.7-point average of the published academic trials, the same letter would have been forecast to bring 17,400 (200,000 × 0.087). Both numbers are real findings about different populations of trials. The distance between them is the reason a nudge is worth trying, and the reason a plan should not be built on one.

Pros

• It is cheap, which is the whole economic case. A letter redesign competes against policies costing orders of magnitude more per person reached.
• It preserves the option a person would otherwise have taken, so being wrong about what people want costs less than it does with a mandate.
• It is testable, and the field's own practice of running controlled trials at scale is what exposed how small the effects are.
• It reaches people who do not respond to prices at all, which is a large share of any population.

Cons

• Average effects at scale are small, and the published literature overstates them by roughly a factor of six for the interventions studied.
• The same tools work with the sign reversed, and most of the choice architecture a household meets was designed by someone with a stake in the outcome.
• Some people object to being steered regardless of the result, and that objection does not show up in a take-up statistic.
• It invites treating a design problem as the whole problem, when the underlying issue is often a price, an income, or an eligibility rule.