Critiques of the heuristics and biases tradition

The chap­ter on judg­ment un­der un­cer­tainty in the (ex­cel­lent) new Oxford Hand­book of Cog­ni­tive Psy­chol­ogy has a handy lit­tle sec­tion on re­cent cri­tiques of the “heuris­tics and bi­ases” tra­di­tion. It also dis­cusses prob­lems with the some­what-com­pet­ing “fast and fru­gal heuris­tics” school of thought, but for now let me just quote the sec­tion on heuris­tics and bi­ases (pp. 608-609):

The heuris­tics and bi­ases pro­gram has been highly in­fluen­tial; how­ever, some have ar­gued that in re­cent years the in­fluence, at least in psy­chol­ogy, has waned (McKen­zie, 2005). This wan­ing has been due in part to pointed cri­tiques of the ap­proach (e.g., Gigeren­zer, 1996). This cri­tique com­prises two main ar­gu­ments: (1) that by fo­cus­ing mainly on co­her­ence stan­dards [e.g. their ra­tio­nal­ity given the sub­ject’s other be­liefs, as con­trasted with cor­re­spon­dence stan­dards hav­ing to do with the real-world ac­cu­racy of a sub­ject’s be­liefs] the ap­proach ig­nores the role played by the en­vi­ron­ment or the con­text in which a judg­ment is made; and (2) that the ex­pla­na­tions of phe­nom­ena via one-word la­bels such as availa­bil­ity, an­chor­ing, and rep­re­sen­ta­tive­ness are vague, in­suffi­cient, and say noth­ing about the pro­cesses un­der­ly­ing judg­ment (see Kah­ne­man, 2003; Kah­ne­man & Tver­sky, 1996 for re­sponses to this cri­tique).

The ac­cu­racy of some of the heuris­tics pro­posed by Tver­sky and Kah­ne­man can be com­pared to cor­re­spon­dence crite­ria (availa­bil­ity and an­chor­ing). Thus, ar­gu­ing that the tra­di­tion only uses the “nar­row norms” (Gigeren­zer, 1996) of co­her­ence crite­ria is not strictly ac­cu­rate (cf. Dun­woody, 2009). Nonethe­less, re­sponses in fa­mous ex­am­ples like the Linda prob­lem can be rein­ter­preted as sen­si­ble rather than er­ro­neous if one uses con­ver­sa­tional or prag­matic norms rather than those de­rived from prob­a­bil­ity the­ory (Hil­ton, 1995). For ex­am­ple, Her­twig, Benz and Krauss (2008) asked par­ti­ci­pants which of the fol­low­ing two state­ments is more prob­a­ble:

[X] The per­centage of ado­les­cent smok­ers in Ger­many de­creases at least 15% from cur­rent lev­els by Septem­ber 1, 2003.

[X&Y] The to­bacco tax in Ger­many is in­creased by 5 cents per cigarette and the per­centage of ado­les­cent smok­ers in Ger­many de­creases at least 15% from cur­rent lev­els by Septem­ber 1, 2003.

Ac­cord­ing to the con­junc­tion rule, [X&Y can­not be more prob­a­ble than X] and yet the ma­jor­ity of par­ti­ci­pants ranked the state­ments in that or­der. How­ever, when sub­se­quently asked to rank or­der four state­ments in or­der of how well each one de­scribed their un­der­stand­ing of X&Y, there was an over­whelming ten­dency to rank state­ments like “X and there­fore Y” or “X and X is the cause for Y” higher than the sim­ple con­junc­tion “X and Y.” More­over, the minor­ity of par­ti­ci­pants who did not com­mit the con­junc­tion fal­lacy in the first judg­ment showed in­ter­nal co­her­ence by rank­ing “X and Y” as best de­scribing their un­der­stand­ing in the sec­ond judg­ment.Th­ese re­sults sug­gest that peo­ple adopt a causal un­der­stand­ing of the state­ments, in essence rank­ing the prob­a­bil­ity of X, given Y as more prob­a­ble than X oc­cur­ring alone. If so, then ar­guably the con­junc­tion “er­ror” is no longer in­cor­rect. (See Moro, 2009 for ex­ten­sive dis­cus­sion of the rea­sons un­der­ly­ing the con­junc­tion fal­lacy, in­clud­ing why “mi­s­un­der­stand­ing” can­not ex­plain all in­stances of the fal­lacy.)

The “vague­ness” ar­gu­ment can be illus­trated by con­sid­er­ing two re­lated phe­nom­ena: the gam­bler’s fal­lacy and the hot-hand (Gigeren­zer & Brighton, 2009). The gam­bler’s fal­lacy is the ten­dency for peo­ple to pre­dict the op­po­site out­come af­ter a run of the same out­come (e.g., pre­dict­ing heads af­ter a run of tails when flip­ping a fair coin); the hot-hand, in con­trast, is the ten­dency to pre­dict a run will con­tinue (e.g., a player mak­ing a shot in bas­ket­ball af­ter a suc­ces­sion of bas­kets; Gilovich, Val­lone, & Tver­sky, 1985). Ay­ton and Fischer (2004) pointed out that al­though these two be­hav­iors are op­po­site—end­ing or con­tin­u­ing runs—they have both been ex­plained via the la­bel “rep­re­sen­ta­tive­ness.” In both cases a faulty con­cept of ran­dom­ness leads peo­ple to ex­pect short sec­tions of a se­quence to be “rep­re­sen­ta­tive” of their gen­er­at­ing pro­cess. In the case of the coin, peo­ple be­lieve (er­ro­neously) that long runs should not oc­cur, so the op­po­site out­come is pre­dicted; for the player, the pres­ence of long runs rules out a ran­dom pro­cess so a con­tinu­a­tion is pre­dicted (Gilovich et al., 1985). The “rep­re­sen­ta­tive­ness” ex­pla­na­tion is there­fore in­com­plete with­out spec­i­fy­ing a pri­ori which of the op­pos­ing prior ex­pec­ta­tions will re­sult. More im­por­tant, rep­re­sen­ta­tive­ness alone does not ex­plain why peo­ple have the mis­con­cep­tion that ran­dom se­quences should ex­hibit lo­cal rep­re­sen­ta­tive­ness when in re­al­ity they do not (Ay­ton & Fischer, 2004).

My thanks to MIRI in­tern Stephen Barnes for tran­scribing this text.