Descriptive norms

Descriptive norms are one of the better-replicated small effects in behavioural science, and the size is what most summaries get wrong. Bergquist, Thiel, Goldberg and van der Linden's second-order meta-analysis of 430 primary studies found social comparison the strongest of six intervention families in field settings (d = 0.37). Read the paper's other numbers too: the unadjusted effect across all families was d = 0.31, publication-bias adjustment brought it to d = 0.18 (about 7 percentage points), and in 32 adequately powered large-scale trials (n > 9,000) it fell to d = 0.04, roughly 2 percentage points. Plan the roadmap against the 2 points. Rhodes, Shulman and McClaran's meta-analysis of norm appeals across 110 articles agrees: consistent support across outcomes, small effects, significant heterogeneity between studies.

The large field trials bear that out. Hallsworth, List, Metcalfe and Vlaev tested UK tax reminder letters on more than 200,000 taxpayers; in the first experiment five test messages raised 23-day payment rates by 2.2 percentage points over the standard reminder. In the second, the specificity gradient is the finding worth stealing: the generic “nine out of ten people pay their tax on time” moved payment by 1.4 points, a norm localised to the recipient's area by 2.2, one localised to their debt type by 3.0, and the combined local-and-debt norm by 5.0 points — a 14.8% relative increase. Descriptive norms as a bundle beat injunctive norms by 1.44 points. Allcott's evaluation of OPOWER's Home Energy Reports, which compare a household's electricity use with its neighbours', found an average reduction of about 2% across roughly 600,000 US households, concentrated in heavy users: 6.3% in the top decile of baseline usage against 0.3% in the bottom. Effects in the earlier Minnesota pilot decayed between letters and rose again when the next one arrived, which is an attention cycle rather than a habit. Allcott later found, across 111 trials covering 8.6 million households, that the first 10 replications substantially overstated efficacy in the next 101 sites, because environmentalist areas and heavy users were targeted first.

The famous demonstrations are much smaller than their reputation. In Goldstein, Cialdini and Griskevicius's hotel-towel experiments, a descriptive-norm card produced 44.1% towel reuse against 35.1% for the standard environmental card; in the second experiment the descriptive norms combined gave 44.5% against 37.2%, and the “guests in this room” provincial norm reached 49.3% against 42.8% for the other norms pooled. Reese, Loew and Steffgen replicated the provincial-norm advantage in two hotels. The authors flag the awkward detail themselves: the sign claimed 75% participation, a figure supplied by the company selling the cards, while observed reuse never exceeded 49.3%. The canonical citation for descriptive norms is also a case study in publishing a norm nobody measured.

The reversal is documented, and it is why this principle needs guardrails. Schultz, Nolan, Cialdini, Goldstein and Griskevicius gave 290 San Marcos households feedback on their electricity use plus their neighbourhood average. Above-average households cut consumption by 1.22 kWh/day (d = 0.55); below-average households raised it by 0.89 kWh/day (d = −0.52) — the boomerang effect. Adding an injunctive cue, a smiling or frowning emoticon, removed it. That remedy is weaker at scale than the textbook version suggests: OPOWER built it in as “Great” and “Good” efficiency labels with smiley faces, and Allcott's regression-discontinuity comparison across that cut-off found no significant effect of the labels, with too little power either to detect an effect or to rule out a small one. Low users in that data did not boomerang; they barely moved. Treat the injunctive add-on as unproven rather than proven.

The nulls are large and specific. Wechsler and colleagues compared 37 US colleges running social-norms marketing campaigns against 61 that were not: no decrease on any of seven alcohol measures, and increases in monthly use and total volume where the programmes ran. Aldoh, Sparks and Harris preregistered a test of dynamic norms — telling people a minority behaviour is growing, the usual advice when your real rate is low, introduced by Sparkman and Walton after it doubled meatless orders in a café — in 1,294 participants, and found no positive effect on attitudes, interest, intentions or self-reported meat consumption at either time point. The surrounding nudge literature has been corrected downwards as well: Mertens, Herberz, Hahnel and Brosch's meta-analysis of 455 effects put the social-reference technique at d = 0.40, but Maier, Bartoš, Stanley, Shanks, Harris and Wagenmakers reanalysed the same data with robust Bayesian meta-analysis and found the decision-information category, the one containing social reference points, dropped to d = 0.00, with moderate-to-strong evidence against a mean effect (BF₀₁ = 33.84). DellaVigna and Linos measured 1.4 percentage points across 126 nudge-unit RCTs covering 23 million individuals, against 8.7 in published academic trials, with selective publication and low power explaining about 70% of the gap.

For interfaces there is one good field experiment, and it splits the funnel in a way the “80% of teams enable 2FA” framing hides. Das, Kramer, Dabbish and Hong showed 50,000 Facebook users one of eight security announcements, seven social-proof variants and one non-social control, promoting Login Notifications, Login Approvals and Trusted Contacts. Stating how many of the viewer's friends used the feature drove 37% more viewers to explore it, but among those who did explore, adoption was no different from the non-social announcement. The norm buys attention, not commitment, so put it on the entry point and fix completion inside the flow. Gerber and Rogers found high-turnout messages beat low-turnout ones in two get-out-the-vote experiments, with descriptive norms moving intention only among infrequent and occasional voters. And Nolan, Schultz, Cialdini, Goldstein and Griskevicius found in a survey of 810 Californians that descriptive normative beliefs predicted conservation behaviour better than any other belief measured while respondents rated them the least important influence on their own decisions, so a preference survey is the wrong instrument for deciding whether to ship a norm.