Modelling Bounded Rationality in Organizations: Progress and Prospects

^†An early draft of this article was prepared as a background note for a workshop on “The psychological foundations of organizational models” held at London Business School, 11–12 May, 2012.

Abstract

Much of the formal modelling work in the organizational sciences relies on Herbert Simon's conception of bounded rationality, and it stakes a claim to drawing on behaviorally plausible assumptions about human behavior and action in organizations. The objectives of our review are threefold. First, we summarize the formal literature by “model families”—classes of models sharing the same analytical structure—to highlight sharply the behavioral assumptions being made. Specifically, we discuss model families involving (a) adaptation through search and learning by individual agents, (b) mutual adjustment of interacting agents, and (c) information aggregation in organizational decision-making. Second, we examine to what extent these models of bounded rationality in organizations are in fact consistent with the behavioral evidence in psychology and other related fields. Finally, we discuss opportunities for further research that strengthens the links between formal modelling in organizations research, and its behavioral foundations. In particular, we highlight the promise of experimental methods that translate organizational models to multiple-subject experiments in the behavioral laboratory.

Acknowledgements

The authors thank Felipe Csaszar, Jerker Denrell, Dan Levinthal, and Thorbjørn Knudsen for helpful comments on earlier drafts of the paper. All remaining errors are ours.

Funding

Puranam acknowledges funding from the European Research Council under grant #241132 for the “The Foundations of Organization design” project.

Notes

^†An early draft of this article was prepared as a background note for a workshop on “The psychological foundations of organizational models” held at London Business School, 11–12 May, 2012.

1. Simon's conception of global rationality was as follows (1997, p. 25):

  This is what I call full or global rationality: people are making their decisions to maximize utility in a world which they either understand exactly or in terms of a known probability distribution (i.e. they are maximizing subjective expected utility).

Since bounded rationality is defined by its departure from global rationality, it follows that there are many possible ways in which rationality can be bounded. There is, of course, a significant body of modeling work with some form of bounded rationality in economics; see, for instance, Rubinstein (1998) or Spiegler (2011), as well as an extensive empirical tradition (e.g. Camerer, 2003; Kahneman & Tversky, 1979). In this literature, bounded rationality is conceptualized variously in terms of the existence of communication and search costs, malleable and unstable preferences, or outright processing errors (e.g. Gilboa & Schmeidler, 2001; Kahneman & Tversky, 1979), while the modeling typically retains the notion of optimization subject to these features.

2. The statement by Schoemaker (1982) in his review of the evidence on expected utility (EU) maximization remains a good summary: “For well-structured repetitive tasks, with important stakes, and well-trained decision makers, EU maximization may well describe the actual decision process.” In other words, unbounded rational choice may be a useful theory when there is a small or well-known space of choices, (expected) utilities associated with the choices have been discovered and are now well known, and there is sufficient difference between the outcomes of choices to make maximization easy. Such situations, Simon argued, are rare within organizations.

3. A process of Selection (through competition) sometimes accompanies organizational models of adaptive rationality (e.g. Levinthal, 1997; Levinthal & Posen, 2007), and may be an alternative (or addition) to the transformation process in terms of generating change over time. Put simply, the individual agents in the model may change their representations through transformation processes, or the distribution of representations in the population may change through selecting out those agents who have poor representations, or both. Since selection through competition does not feature in all models of adaptive rationality, we only discuss it when relevant.

4. Models with the first feature but not the second also exist in the organizational literature, and often invoke maximization within an imperfect representation (e.g. Gulati & Puranam, 2009; Kretschmer & Puranam, 2008; Puranam & Vanneste, 2009; Siggelkow, 2002). Since there is considerable variation in the nature of the task environments and representations in these papers, we do not attempt to summarize these here.

5. The unitary actor assumption can be justified if there are rules of preference/belief aggregation, and selection into or out of an organization does not destabilize such processes; if an organization is effectively controlled by a powerful individual, or a governing coalition which has stable preferences, beliefs and choice procedures (see, for instance, Hug (1999) for a formal analysis of the conditions under which unitary actor assumptions may be more or less plausible).

6. The basic bandit setup usually assumes a stable task environment, that is, the probability distribution of payoffs for each arm is fixed. However, the model can be easily amended to also consider dynamic task environments where the probability distributions fluctuate over time to depict changing markets or technological conditions for the opportunities that each alternative represents (Posen & Levinthal, 2012; Stieglitz et al., 2014). Similarly, competence development and lock-in to particular alternatives may be represented by making the probability distribution conditional on the number of times an alternative has been chosen (Denrell & March, 2001).

7. A moderate level of exploration is assumed. At very high levels of exploration, the effects of learning are suppressed, since choices do not discriminate among alternatives, regardless of what has been learned about them.

8. We mean similarity beyond that induced by the language of description. For instance, if we want to compare individual and organizational screening functions, we must look for similarities beyond those generic to any screening function.

9. Field experiments would be even stronger if feasible to set up in a way to allow for precise tests of mechanisms (e.g. Bloom et al., 2013).