Availability heuristic

The strand of this literature with the best record is availability-by-recall: judging a risk by counting the cases you have actually encountered. Hertwig, Pachur and Kurzenhäuser tested four candidate mechanisms against people's choices in three risk environments, and availability-by-recall and a regressed-frequency model conformed best. Pachur, Hertwig and Steinmann pitted availability against the affect heuristic across judged frequency, value of a statistical life and perceived risk: availability-by-recall, defined as occurrences of the risk in the respondent's own social network, fitted best overall, and encounters with risks in the media played a negligible role. Efendić's experiment supplies the causal version — asking participants to think of more deaths in their own social network raised both mortality judgments and the value placed on a single life, irrespective of affect, media coverage of the risk, or retrieval time. The availability doing the work is counted lived instances, not exposure on a screen, and a screen is the only channel an interface has.

The famous demonstrations are weaker than their reputation. Lichtenstein, Slovic, Fischhoff, Layman and Combs found a biased subjective frequency scale for causes of death — small frequencies overestimated, large ones underestimated, specific causes exaggerated or discounted — which participants could not correct when instructed to try. But that first pattern is exactly what a regressed-frequency model predicts with no appeal to memory, and in Hertwig and colleagues' comparison the two mechanisms performed at a nearly indistinguishable level. Sedlmeier, Hertwig and Gigerenzer challenged the letter-frequency task head on: across three studies, judgments of whether a letter is more common in first or second position generally followed the actual proportions, with low values overestimated and high ones underestimated. You cannot assume a user's frequency estimate is memory-distorted just because it came from memory.

The second mechanism is the one that inverts the obvious advice. Schwarz, Bless, Strack, Klumpp, Rittenauer-Schatka and Simons had people recall either six or twelve of their own assertive behaviours and then rate their assertiveness; those who produced twelve rated themselves less assertive, because twelve was hard to come up with and the difficulty itself was read as evidence. Weingarten and Hutchinson meta-analysed the paradigm across 142 papers, 263 studies and 582 effect sizes: the standard few-versus-many manipulation shows a medium-sized average effect, between about a third and a half of it mediated by subjective ease, with publication bias that reduces the standard effect by up to a third. Manipulations departing from that standard design produce smaller and often reversed effects, less strongly mediated by ease, so the direction of a retrieval-difficulty manipulation in a new flow is not predictable from the literature.

Both of the big clean tests come up empty. Yeager, Krosnick, Visser, Holbrook and Tahk ran seven highly cited social-psychological phenomena on randomly sampled, representative US adults; six replicated, more weakly than the meta-analytic estimates from student samples, and the one that did not was the ease-of-retrieval cross-over interaction. Groncki, Beaudry and Sauer found no ease-of-retrieval effect on eyewitness confidence or related judgments across three studies, then failed to replicate Schwarz et al.'s original result directly (N = 661), with both frequentist and Bayesian tests showing no evidence of an effect. Note what did work in three of their four studies: the manipulation moved perceived task difficulty as intended, and the judgment still did not move. A difficulty manipulation that demonstrably lands is not sufficient to move the number a user reports.

The one applied field result points the same way as Schwarz, not as the folk version. Fox asked business students mid-course to list either 2 or 10 ways the course could be improved. The ten-slot group rated the course higher — 5.52 against 4.92 on a 1–7 scale — while producing no more criticisms (2.1 against 1.6). What predicted the rating was the ratio of criticisms produced to criticisms solicited, not the count. Asking for more examples of a problem made the problem look smaller, which makes the cap you put on a feedback prompt a design decision with a measured sign.

There is, as far as this review found, no controlled evidence that making examples easy to recall works as a product intervention: no experiment tests whether a case-study shelf, a customer-story carousel or a “people like you” module changes a user's estimate of anything. The nearest controlled work shows the mechanism operating against accuracy. Fagerlin, Wang and Ubel had people choose between bypass surgery and angioplasty from cure rates given in prose plus patient anecdotes: 41% chose bypass when the anecdotes were representative of the cure rates and 20% when they were not. Reinforcing the same statistics with a pictograph and a quiz removed the anecdotes' influence (38% against 44%); the quiz alone did nothing (27% against 28%) and the pictograph alone was what worked. Tversky and Koehler's support theory covers the reading side — unpacking a category into explicitly named components raises its judged probability — so a list of named failure modes beside an offer is itself a probability manipulation. Build exemplars as a device that moves estimates in a direction you have not measured, and keep a graphical base rate on the same screen.