Social comparison

Social proof and social comparison read the same raw material for different ends: social proof treats other people's behaviour as evidence about the right action under uncertainty — reviews, “nine out of ten people pay on time” — while social comparison treats their standing as evidence about the self, in a rank, a benchmark or a percentile. Descriptive norms belong with the first. Festinger's 1954 paper is theory, not experiment, and the hypotheses that read as design instructions are the ones usually skipped. Hypothesis II is conditional — “to the extent that objective, non-social means are not available, people evaluate their opinions and abilities by comparison respectively with the opinions and abilities of others” — so an objective standard of your own lowers the pressure to compare. Hypothesis III says the tendency to compare falls as the ability gap widens, with the person oriented instead toward a point “slightly better than his own performance,” which rules a distant leader out as a comparison target. Hypothesis IV posits a unidirectional drive upward for abilities, largely absent for opinions. Hypothesis VI says people stop comparing when continued comparison implies unpleasant consequences, with hostility or derogation attached — in a product, churn rather than effort.

The claim that comparison produces motivation does not survive the meta-analysis. Gerber, Wheeler and Suls pooled 60-plus years of the literature and found contrast, not inspiration, to be by far the dominant reaction: when participants' standing was varied, estimates ran −0.75 to −0.65 for ability and −0.83 to −0.65 for affect. People still choose to compare upward when offered an up-or-down option, with no evidence that downward comparison becomes dominant even under threat. Where a no-comparison control existed, upward and downward targets produced comparable, contrastive reactions. Assimilation — the “they inspire me” story — showed up only as very weak effects under similarity priming, dependent on which publication-bias estimator was used. Motivation is a conditional outcome you have to engineer.

Ranking changes behavior, and the sign depends on the setting. In a UK fruit farm, Bandiera, Barankay and Rasul found that rank incentives for teams cut average productivity by 14% while a monetary-prize tournament raised it by 24%; the losses came only from teams at the bottom of the productivity distribution and the tournament gains only from those at the top. Gill, Kissová, Lee and Prowse randomly allocated rank among equal-effort subjects and found the response U-shaped — “first-place loving and last-place loathing” — so in a short task the bottom pushes back rather than giving up. Rank also sticks: Murphy and Weinhardt, across English school students, found primary-school class rank affected secondary-school test scores, confidence and subject choice independent of ability, more so for boys, and under new peers and teachers unaware of the old ranking. Megalokonomou and Zhang, using random classroom assignment in China, found higher assigned rank raised later performance, more where rank was salient. None of this tells you which sign your own ranking will take.

Ranking against named people and benchmarking against an average are not one pattern. Azmat and Iriberri gave subjects the session average alongside their own score: performance was 17% higher under piece-rate pay and unaffected under flat-rate pay, regardless of relative position — a benchmark moves effort only where effort pays. The affective cost landed anyway, with the happiness gap between those above and below average widening by 8 percentage points under piece rate. Schultz, Nolan, Cialdini, Goldstein and Griskevicius found the reversal in the field: a neighbourhood-average energy message produced savings among heavy users and a boomerang — increased consumption — among households already below the average, and an injunctive cue conveying approval or disapproval eliminated the boomerang. Hermes, Huschens, Rothlauf and Schunk scored individual improvement rather than standing; in a class-randomised trial with primary-school children that raised motivation, effort and maths performance for low achievers without hurting high achievers, with the gains concentrated in girls.

The product-shaped evidence is thinner than the volume of leaderboards shipped would suggest. Sailer and Homner's meta-analysis of gamification in learning found small significant effects — g = 0.49 for cognitive, 0.36 for motivational and 0.25 for behavioral outcomes — with the cognitive effect stable among methodologically rigorous studies and the other two less so, and the effective ingredient for behavioral outcomes was competition combined with collaboration, not competition alone. The best field test is Patel and colleagues' STEP UP trial of gamified step goals over 24 weeks with a 12-week follow-up: competition, support and collaboration all beat control during the intervention — by 920, 689 and 637 daily steps — and only competition was still significantly above control at follow-up, by 569. That is one trial, in a health context, with a four-person comparison group rather than a public ladder.

On the social-media side, the kind of evidence matters. Yoon, Kleinman, Mertz and Brannick found depression related to time on social network sites and to checking frequency with small effects, and more strongly to general and upward comparison — all cross-sectional. Verduyn and colleagues conclude that comparison on these sites is typically contrastive and upward and tracks lower well-being, with passive use raising comparison most. The direction is not settled: in Carraturo and colleagues' review, the only two studies able to establish direction found depression predicting comparison and envy rather than following from it. The causal estimates are much smaller: in Allcott and colleagues' randomised deactivation experiments before the 2020 US election, six weeks off Facebook improved an index of happiness, depression and anxiety by 0.060 standard deviations and six weeks off Instagram by 0.041. Treat comparison as a real mechanism with a small measured causal footprint, and any leaderboard of your own as untested until you have run it.