Issues with boundedly rational models

A first issue with boundedly rational models is that research on heuristic strategies has shown that identifying the specific decision strategy used is problematic. It may be true, first of all, that boundedly rational decision making converges with optimisation (Doucouliagos, 1994), validating Friedman’s contention that behaviour can be modelled as if it is optimising. However, whether the two types of decision making converge is an empirical question that argues for more study of bounded rationality, rather than dismissing it as unnecessary (Conlisk, 1996). Mathematical models for this purpose have been developed (Rubinstein, 1998), but have also been criticised for being armchair models without enough basis in psychology, decision theory, and empirical evidence (Friedman, 1998; see also Simon's chapter in Rubinstein, 1998).

Several empirical studies have investigated the use of heuristics, and have generally found it difficult to identify the specific decision protocol used. In the experiment discussed above that entailed choosing the company with the highest profit (Gigerenzer et al., in press; Rieskamp & Hoffrage, 1999), subjects were asked to select the best company from a set of four

strategies were being used because respondents’ choices could have arisen either from integrative or lexicographic strategies, and both theories fit the data. In the German city problem described above, the different decision protocols led to the same answer in 92% of the pairwise comparisons (Broder, 2000), making it impossible to identify the simulated protocol just from the choices made. The identification problem is further exacerbated when considering choice probabilities of an entire sample: if some respondents make choices using integrative protocols while other use heuristic protocols, the overall sample probabilities can still be compatible with RUM theory (Koning & Ridder, 2003). It would seem from the perspective of the whole sample that decision-making was integrative, making it difficult to identify specific individuals’ heuristic strategies.

These results are an example of a ‘flat maximum’ (Broder, 2000), and are not confined to difficulties identifying the use of heuristic strategies. Linear, integrative models continue to perform well even after the parameter weights are changed, as long as the signs of the

parameters are maintained (Broder, 2000). For this reason, Payne & Bettman (2001) modelled both a standard decision protocol that attached weights to different attributes, but also

modelled an ‘equal weight strategy’ in which all attributes were equally weighted. While this is a problem for identifying the decision protocol that respondents use – how is the researcher to identify the correct protocol when several fit the data? – it is also an argument in favour of bounded rationality. If a number of different decision protocols can all lead to the same alternative in a given decision environment, then a cognitively simple, heuristic strategy is more efficient than a holistic, integrative strategy (Gigerenzer & Selten, 2001b).

A second criticism of models of bounded rationality is their limited applicability. Nearly all research using EBA and fast and frugal heuristics has assessed their validity using binary attributes, so the models need to be expanded in order to apply more generally in

binary attributes that demonstrate the difficulties. Rotondo’s (1986) nested model of EBA that allowed prices in the choice set to take several values, shows that the number of nests would expand exponentially with the number of additional levels modelled. Expanding such a model to include several multi-level attributes would create a cumbersome number of nests.

The number of alternatives in the choice situation is also an issue. Rieskamp & Hoffrage (1999) did expand fast and frugal heuristics to situations of more than two alternatives. However, subjects were taught the relative importance of choice attributes before engaging in decision-making (Broder, 2000). They were thus all using the same set of attribute weights. In the case of GM food, consumer research suggests that consumers will not all place the same weight on the attributes of food. The choice data needs to be analysed both for the decision protocol used and the weights given different attributes.

Because each decision rule has limited applicability, bounded rationality has been accused of being ad hoc (Conlisk, 1996). At root, this criticism resembles the infinite regress problem. If bounded rationality seems ad hoc, this is because it fits the decision protocol to the data but does not say how the protocol is chosen. If bounded rationality were to propose an invariable rule about how decisions were made or an invariable rule about how decision protocols were chosen, then it could no longer be accused of being ad hoc. The same rule would apply in all situations. This would end the infinite regress of deciding how to decide how to decide, etc. The infinite regress problem was discussed above. In essence, it is a problem with boundedly rational models and one that cannot be resolved. However, as also discussed above,

maximising models have problems with infinite regress, too, so the issue is not limited to boundedly rational models.

The perception that boundedly rational models have limited applicability has also led this area of economic research to be accused of unnecessarily multiplying the number of options that

must be considered (Rabin, 2002). It is true that a number of decision protocols have been suggested. However, only a few have been widely discussed and intensively researched, so the set of standard boundedly rational protocols is quite small. Furthermore, given the complexity of advanced discrete choice modelling, it would be difficult to argue that researchers are interested in reducing the number of parameters to be estimated, or that economic modellers are averse to complexity (Rabin, 2002).