Regret aversion

Bell (1982) and Loomes and Sugden (1982) formalised the same idea independently: an outcome's value is the outcome set against what the rejected alternative would have produced. That counterfactual separates this principle from the two it is confused with. Loss aversion needs only a reference point, something you hold that an action would take away; status quo bias needs only inaction to be available. Regret needs a rejected option whose fate can be compared with yours, so it turns on information rather than on ownership or inertia. Whether your product later tells the user what the other path would have given them decides whether this principle applies at all.

That switch has been tested directly. Zeelenberg, Beattie, van der Pligt and de Vries (1996) gave 78 undergraduates a safe and a risky gamble matched for attractiveness, then varied only which gamble's outcome they would be told afterwards. Participants were more risk-seeking when feedback was promised on the risky gamble and more risk-averse when it was promised on the safe one, with the choice-only condition in between; a third experiment with real money reproduced the direction. People choose the option that spares them from learning what they missed, so a screen that reveals the forgone outcome pushes behaviour toward the revealed branch.

That paradigm has been picked apart by its own builders, and Bleichrodt and Wakker's (2015) retrospective records both halves. Loomes and colleagues found the preference cycles regret theory predicts far more common than the reverse, but Day and Loomes (2010) found the cycle the theory forbids prevailing, several independent groups report “similarity cycles” it cannot produce, and Harless (1992) concluded that regret effects are primarily framing effects that “occur only when the decision is framed in a way that sharply directs the decision maker to compare acts and states”. Starmer and Sugden (1993) landed the sharpest blow: much of the earlier support turned out to be an event-splitting effect, present in the two problems where splitting also predicted it and absent in the two where it did not. That puts the effect in explicit side-by-side comparison — what a plan grid or a checkout summary is, and a paragraph of marketing copy is not.

The correlational base is large and independent of that paradigm. Sandberg and Conner (2008) meta-analysed anticipated regret as an addition to the theory of planned behaviour: r+ = .47 with intention across 25 studies and 11,254 people, r+ = .28 with later behaviour, and a contribution independent of the model's existing predictors. Brewer, DeFrank and Gilkey (2016) pooled 81 studies, 45,618 participants and 128 effect sizes and landed in the same place, r+ = .50 with intention and r+ = .29 with behaviour. The signs split by direction, which is what matters for a build: inaction regret went with stronger intentions (r+ = .52) and more behaviour (r+ = .29), action regret with weaker intentions (r+ = −.45) and less behaviour (r+ = −.28). Their own limitation section is blunt — 58% of the studies were cross-sectional, 74% used convenience samples, and “the correlational designs widely used in the studies precludes strong causal inference”. Their pilot work also suggested direct regret appeals might elicit reactance, so they put the construct in a quote from a local parent rather than aiming it at the reader.

Prompting regret deliberately makes people buy sooner, not hesitate. Simonson (1992) ran 218 undergraduates through a purchase-timing scenario: buy the wedding gift in the July sale or wait for August, at a store with no returns on sale items. Asked first which mistake would upset them more, 63% chose July against 54% in control, and the better-known, higher-priced brand took 67% of choices against 50%. Two null controls matter as much: telling people they would learn which option had been better, and telling them their decisions would be rated, moved nothing significant. The lever is anticipating the feeling, and it points at immediate purchase and at the safe brand — in hypothetical student scenarios, so the direction is instructive and the sizes are not.

The field record is thinner. Sandberg and Conner (2009) sent 4,277 women a cervical screening invitation; attendance was 21% in control, 26% among those who also got a theory-of-planned-behaviour questionnaire and 26% among those who got that questionnaire plus anticipated regret items. The regret items added nothing over asking the questions at all, and the quoted 65% against 44% advantage exists only inside the self-selected subset who returned their questionnaire, and only among strong intenders. Haisley, Volpp, Pellathy and Loewenstein (2012) got a larger result from structure rather than wording: across 1,299 employees in 14 offices, health risk assessment completion was 64% under a team lottery where non-participants learn whether they would have won, against 44% for a $50 grocery certificate and 40% for the existing $25 payment. That design confounds regret with peer pressure and a different enrolment procedure on purpose.

Regret aversion earns its place in a product on the other side: lowering what the user stands to regret. Janakiraman, Syrdal and Freling (2016) meta-analysed 21 papers on return policy leniency: leniency raises purchases more than it raises returns, and the five dimensions pull differently, with money and effort leniency increasing purchases, scope leniency increasing returns, time and exchange leniency reducing them. Suwelack, Hogreve and Hoyer (2011) found a money-back guarantee raises purchase intention and willingness to pay a price premium, and that guarantees on experience goods need stricter return conditions than on search goods. Petersen and Kumar (2015) ran the field version with 26,000 customers over six months: a satisfactory return experience lowers perceived risk on current and future purchases rather than being a pure cost. Lowering the stakes also changes who walks in: trial-acquired customers had average customer lifetime value 59% lower than regular customers, while responding more strongly to marketing and to usage (Datta, Foubert and Van Heerde 2015). A trial needs its own cohort, not the same forecast.

The action-versus-inaction story that usually accompanies this principle is its weakest part. Spranca, Minsk and Baron (1991) found harmful omissions rated less immoral than harmful commissions, with no mirror-image effect for good outcomes. But the famous claim that we regret inactions more over a lifetime rests on Gilovich and Medvec (1994): 84% of 32 adults named an inaction as their bigger long-term regret. Richardson and Gilovich (2023) re-ran it with 2,600 usable cards: the short-term pattern replicated, 58% naming an action against 42% an inaction, and the long-term result did not, 51% action against 49% inaction, p = .35. Roese and Summerville (2005) had already reframed the pattern as opportunity rather than inaction — pooling 11 datasets and 4,054 regrets, education was the top regret at 32.2%, and high-opportunity life domains produced more intense regrets than low-opportunity ones (d = .21), while framing manipulations at the point of recall did nothing consistent (d = −.02). Regret follows the door that is still open better than it follows inaction. The record overall: two meta-analyses agreeing on direction and size, one clean experiment on the moderator that matters for software, a contested lab paradigm, no causal evidence from a product interface, and one evidenced lever that is structural — returns, guarantees, reversibility — not regret-framed copy.