Use Norms as a Conditional Tool, Not a Button
The central question is whether social-norm messaging deserves to be treated as a reliable behavior-change tool. The strongest answer supported by the evidence is no—not as a general button—but also not a blanket dismissal. Norm appeals can alter perceived prevalence or approval and can sometimes shift an immediate choice. Their reliability depends on the reference group, the setting, the practical alternatives available to the person, the endpoint being measured, and the quality of the evidence behind the claim. A sound decision is therefore conditional: test the norm appeal where its mechanism fits, keep its contribution separate from the rest of the intervention, and scale only what survives honest comparison and follow-up.
What “not reliable” means
The broad randomized evidence sets the first boundary. A preregistered systematic review and meta-analysis examined 89 randomized studies from 60 papers, involving 85,759 participants aged 16 or older in developed countries. The studies targeted health behaviors such as diet, screening, vaccination, alcohol use, prescribing, physical activity, and hand hygiene. In a conventional random-effects analysis, the average effect of social-norm messaging was small but statistically positive: Cohen’s d was 0.14, with a 95% confidence interval from 0.09 to 0.19. That is an estimate of an average across the included studies, not a promise about the next clinic, campaign, or population.
The distribution beneath that average was wide. Some studies reported moderate or large positive effects, many reported trivial effects, and some reported small negative effects. The estimated prediction interval ran from d = −0.43 to 1.19. A narrow confidence interval around the average could coexist with that broad range because the two intervals answer different questions. The first concerns uncertainty about the mean under the model; the second reflects how much a new setting might differ from that mean. The intervention could therefore be detectable as slightly positive in the aggregate while remaining difficult to predict in a particular application.
The evidence became weaker when the review examined how the literature may have been selected. Positive effects were overrepresented among studies with low precision, a pattern consistent with publication bias. A Bayesian analysis without bias adjustment reproduced the small positive estimate. Robust Bayesian meta-analysis that adjusted for potential publication bias reduced the estimate to d = 0.01, with a 95% credible interval from 0.00 to 0.09. The review reports strong evidence for the null in that adjusted model and very strong evidence for publication bias. The exact model and priors matter, and the adjustment does not prove that every positive study is spurious. It does mean that the published small average is not a dependable foundation for claiming a general health-behavior effect.
This is the difference between evidence of absence and absence of evidence. The adjusted analysis is evidence against a reliable average effect within the review’s scope. It is not evidence that no norm appeal can ever work. The study-level variation remains real, and a useful local effect may be hidden inside a near-zero average. What the broad evidence removes is the shortcut from “some interventions produced positive results” to “norm messaging is a trustworthy general-purpose lever.”
The moderator analysis removes another shortcut. The review tested message type, delivery mode, health domain, and target population. Frequentist omnibus tests showed variation across categories, but no individual category reliably differed from the grand mean in the fuller Bayesian and bias-adjusted analyses. A small positive signal among university students disappeared after adjustment. The results did not identify a universally superior descriptive, injunctive, digital, physical, clinical, or population-targeted format. This does not show that design details never matter. It shows that broad labels did not explain enough of the variation to function as a recipe.
The reason for that limit is visible in the intervention itself. “Digital delivery” can describe a carefully timed reminder or a message encountered while someone is distracted. “Students” can describe people who already accept the target norm or people who see the reference group as irrelevant. “Diet” can include a low-friction food choice or a behavior requiring money, planning, and repeated access. The categories combine different mechanisms and constraints. A null result at the category level can coexist with a meaningful condition at the level of one audience, one behavior, and one moment.
The field evidence supplies such conditional cases. Five experiments tested a humorous “Be like Toni” message family that combined descriptive and injunctive cues. The intervention reduced paper-towel use in a university restroom and replicated in a more diverse Christmas-market restroom. Prescriptive and proscriptive versions both reduced use against a control in one university setting, without a meaningful difference between the wordings. In a bakery, the intervention increased voluntary mask wearing, with a reported odds ratio of 2.03 and effect size d = 0.39. The same message family did not change plastic-lid use at a train-station coffee shop: the effect was d = 0.02, with p = 0.881. The authors summarize the five experiments as showing effects in four cases and a null in one.
These field results do not overturn the adjusted meta-analysis. They answer a different question with direct observation in particular German settings.