Independence and interdependence: Lessons from the hive

Background

At the end of spring or the beginning of summer, a honeybee colony that has grown too large tends to split. The queen bee leaves with approximately two-thirds of the worker bees, while a daughter queen stays in the maternal nest with the others. To survive, the bees that have left must quickly find a new nest. Empirical research has shown that they do so by means of a striking decision process (e.g., Lindauer, 1955; Seeley, 2010; Seeley et al., 2006). This involves a “search committee” of several hundred “scout bees” who roam the surrounding area in search of potential nest sites and then return to the swarm to draw the others’ attention to any good sites they have discovered.

In particular, after discovering a potential nest site, each scout bee performs a waggle dance whose orientation encodes the site’s location and whose duration encodes her assessment of the site’s quality. The better she perceives the site to be, the longer she dances. At first, the scout bees rely on discovering potential nest sites by chance, but once they observe other scouts dancing, they are more likely to investigate the sites advertised by those others. If they agree with the positive assessment of a site, they join the dance for it. In this way, sites supported by dancing bees are visited and inspected more often than other sites and, if not supported in error, tend to receive even more support. The process leads to a “consensus” relatively quickly – in one or two days – when the support rallies around one popular site. Once a critical threshold is reached, the swarm moves there. Crucially, when there are quality differences between different potential nest sites, the bees usually find one of the best ones (Seeley and Buhrman, 2001).

There are three constraints that make the speed and accuracy of this decision process all the more surprising. Firstly, the agenda of options is not straightforwardly given, unlike in the decision problems to which Condorcet’s jury theorem is usually applied. An indefinite number of places in the bees’ environment could in principle become candidates for nest sites, and a suitable method of agenda setting is needed to sort out the serious options from the non-starters. Secondly, although individual bees have some remarkable capacities, they are still fairly simple organisms, and a simultaneous and comparative assessment of all potential nest sites is beyond any bee’s capacities. For this reason, the bees’ collective decision process must not place high cognitive demands on any individual bee. Thirdly and relatedly, the bees’ assessment of potential nest sites and their communication are subject to a significant amount of noise, and the decision process must therefore be error-tolerant. So which features of the bees’ decision process explain its remarkable speed and accuracy, in light of these constraints? A stylized model of the process helps to reveal what drives its success. (An explanation of the underlying formal model is provided in the Appendix; here we merely indicate the main features.)

A simple model

The bees, like humans in a Condorcetian jury decision, face an epistemic collective decision problem, but, unlike in an ordinary jury decision, the set of options is more open-ended. So there is a group of individuals (in the present case several hundred scout bees) that has to make a choice among a number of options (different possible nest sites) but that number can be very large. As before, the individuals have common fundamental preferences but possibly different information. We can represent these preferences by assuming that each possible nest site has an objective, though unknown, quality level. The decision process extends over multiple time periods (where time, for simplicity, is discrete), and we model each individual scout bee’s behavior over time (employing the formal model in List et al., 2009, described in the Appendix). Technically, the model is an “agent-based” model, which is defined by specifying, firstly, what state each individual scout bee can be in during each time period and, secondly, how each scout bee changes her state from one time period to the next.

Let us begin by considering each scout bee’s possible states. In any given time period, a scout bee can be in one of two states.

We next specify how a scout bee changes her state from one time period to the next. There are two cases to consider.

Having specified the possible states in which each scout bee can be, as well as the way each bee changes her state from one time period to the next, we can use computer simulations to see how the states of a collection of scout bees change over time and how long it takes for a “consensus” – defined as a sufficiently large plurality of support – to emerge for a particular nest site. The computer simulation can be started by assuming that in the first time period all bees are in the non-dancing state. For present purposes, it suffices to summarize the findings in qualitative terms (for detailed quantitative results, see List et al., 2009 and the summary in the Appendix below).

The determinants of the bees’ collective performance

As noted, the bees’ collective performance depends on each scout bee’s probability of finding and visiting one of the different potential nest sites and her competence and independence in assessing any such site once it has come to her attention. Let us compare different assumptions about each of these determinants of the process:

Conditions favoring the use of the bees’ decision-making mechanism

Although this decision-making mechanism has evolved in honeybees choosing nest sites, the structural features that make it work are transferable to other multi-agent systems. Any collection of agents that has to make fast and accurate decisions without an exogenously defined agenda can in principle implement the bees’ decision-making protocol, as formally captured by the model we have described. What is needed is the ability to roam the space of possible options, to identify and independently rate potential options, however fallibly, and to draw each other’s attention to options that are worth investigating. We have seen that organisms as simple as individual bees have this threefold ability, but nothing in this package of skills is tied to a particular species, a particular decision problem, or a particular biological realization. We can view the bees’ decision-making mechanism through a purely functionalist lens, abstracting away from the case of the bees, and ask in which decision-making environments a functionally similar mechanism would be useful.

In epistemic decision problems, three conditions seem to favor the use of such a mechanism: (1) an open-ended agenda; (2) (relatively) high stakes; and (3) (relatively) high opportunity costs of indecision. In the bees’ case, as we have seen, all three conditions are clearly present. Firstly, the agenda of potential nest sites is not straightforwardly given. Secondly, whether they find a good nest site, a mediocre site, or only a bad one can affect their survival and reproductive success. Thirdly, indecision is not an option, since failing to reach a timely decision can threaten the swarm’s survival.

Generally, we suggest that the less exogenously well-demarcated the agenda, the higher the stakes, understood as the utility differences between the options, and the higher the opportunity costs of indecision, the more a group can benefit from applying the bees’ mechanism. How successful a group will be if it does so then depends on its ability to balance interdependence in signaling to each other which options are worth considering with independence in the individual assessments of those options, over and above each individual’s epistemic competence.

Agenda-setting on the Supreme Court

If there is any governmental institution that routinely faces epistemic decision problems, it is the judiciary. Apart from a relatively small number of political cases, in which fundamental preferences differ along conservative and liberal lines, most cases present issues in which the judges have similar preferences but, at most, differing beliefs. This is true even at the level of the Supreme Court. Despite the strong selection pressure for hard cases to appear at the higher levels of the judicial hierarchy – easy cases are more likely to be settled, or never to be appealed – 47% of the Court’s decisions were unanimous in the 2009 Term, ⁹ and in the cases with dissents, some large fraction involved disagreements about facts or diverging predictions about the consequences of a ruling one way or another. At the agenda-setting stage, a recent study finds that “legal considerations strongly influence justices’ agenda-setting behavior”, although ideological differences play a role as well (Black and Owens, 2009: 1063). We bracket the latter point and assume, without too much distortion of reality, that the justices share common fundamental preferences about what types of cases to hear.

The Supreme Court is, however, unique in the broad control it enjoys over its own agenda. Since the Judiciary Act of 1925, the Court takes almost all of its cases by granting a petition for certiorari, and the Court possesses extremely broad discretion in deciding which petitions to grant. As a matter of practice, the Court is most likely to grant cases in which the lower courts have disagreed, or in which a federal statute has been invalidated on constitutional grounds. ¹⁰ Yet in the end these practices are just rules of thumb or guidelines, which the Court follows or ignores according to circumstances. The Court’s discretion is increased by the sheer number of certiorari petitions that flood it every year – over 8000 on average in recent years, ¹¹ of which the Court usually grants slightly over 80 on average. ¹² The problem for the Court as a body is to sift through the enormous mass of petitions to find the 1% that should be given the Court’s full attention.

The Court’s decision-making environment, in other words, combines nearly unlimited agenda control and high search costs. A great deal of the Court’s business lies in deciding what to decide (Perry, 1991). In this environment, the Court has developed, over time, a set of elaborate procedures for deciding what cases to take. The basic norm is a Rule of Four – the votes of four of nine Justices suffice to grant a certiorari petition for a full hearing. The voting takes place on petitions that are placed on a “discuss list” for the Justices’ weekly conferences; any Justice may place petitions on the list.

Yet how can the Justices sort through, in a single term, more than 8000 certiorari petitions to decide which ones even to place on the list at all? To do so individually would consume most of the Justices’ time. Accordingly, most Justices – at the time of this writing, all but Justice Alito – participate in an institution that seeks to generate economies of scale at the preliminary stage of identifying plausible candidates for the discuss list. This institution is called the “cert[iorari] pool.” The law clerks (the scout bees) employed by the participating Justices divide up the mass of petitions among themselves and then circulate a memorandum to all the Justices in the cert pool. In some chambers, a clerk for Justice X will prepare a second memorandum, but often this is done, or done well, only for petitions that the initial writer of the pool memo has recommended to be granted. The pool memo writer, in other words, has some de facto leeway to shape the Court’s docket, especially by recommending denial. The institutional pressure to “deny cert” is enormous, and a recommendation of denial is rarely contested or closely scrutinized by clerks in other chambers, unless a case has obvious political import.

This institutionalized process of search-and-agenda-setting bears an imperfect, but illuminating, resemblance to the honeybees’ decision procedure. The cert pool is like the bees’ scout committee. Just as interdependence among the bees means that bees are more likely to inspect nest sites advertised by others, so too the effect of the cert pool is that Justices and clerks are more likely to pay close attention to petitions advertised by pool clerks as “certworthy,” or good candidates for a grant. Yet the evaluation of plausibly certworthy candidates is largely independent, both among the bees and among the clerks and their Justices. Once the cert pool writer has identified a plausible candidate for the discuss list, Justices and clerks in other chambers independently evaluate the petition to decide whether to place it on the Court’s agenda. Once a petition has been granted and a case has been given a full hearing, moreover, each Justice independently evaluates the legal claims. The Court’s process combines interdependence at a crucial preliminary stage – identifying, from the mass of petitions, plausible candidates for the agenda – with independence at all later stages of the decision-making process.

How well does the Court’s process work? On one level it is impossible to know, because we have no independent benchmark assessment that would tell us which petitions in fact warrant a full hearing. Yet we can offer some conditional conclusions, and one implication. If and to the extent that the Court’s decision-making environment at the certiorari stage is understood as having the same features as that of the honeybees – an open-ended agenda, real stakes, and real opportunity costs of inaction – then the Court’s process is well-engineered for that environment, subject to some improvements we describe below. Whether those conditions are met is a matter for debate. Many critics of the Court believe that the Court should hear more cases and thus should grant more petitions; presumably these critics believe that it is better for the Court to make more decisions than fewer, perhaps because decisions by the Court clarify the law and promote legal, economic, and political certainty. Views such as these implicitly suppose, in other words, that there are high opportunity costs, from the social point of view, if the Court too often fails to reach consensus on which cases to hear. In this light, interdependence at the early stages of the Court’s agenda-setting process is desirable; perhaps even more interdependence than currently exists would be desirable. Conversely, however, the marked independence of the later stages of the process is desirable to the extent one thinks that the marginal stakes in the Court’s decisions are high, so that information cascades and other phenomena associated with the lack of independence are especially harmful when they cause the Justices to reach consensus on the wrong (or worse) answer.

Suppose one believes that the Court’s decision-making environment does present the combination of factors that make the honeybees’ interdependence–independence mechanism useful. An implication is that the cert pool should incorporate most but not all of the Justices, to optimize the balance of interdependence and independence. Recall that although the honeybees preferably assess nest sites advertised by others, they nonetheless each retain a nonzero probability of stumbling upon potential nest sites on their own. This residual independence, even at the agenda-setting stage, is crucial to the efficiency of the mechanism. As noted above, where interdependence reaches its limiting maximal value, “there is not enough noise in the system for bees to discover any new sites not advertised by others. Small noisy deviations from perfect [interdependence] are necessary to permit the discovery of new sites” (List et al., 2009: 758). Likewise, a cert pool containing all Justices and their clerks would in effect place in the hands of a single 20-something law clerk a real measure of de facto power to set the Court’s agenda, perhaps by burying certworthy cases. The existence of Justices who review all petitions independently of the pool is beneficial for the group, as it provides an independent check on the work of the pool clerks and creates a small amount of beneficial noise in the system, as the nonpool Justices and their clerks search for certworthy cases in parallel to the official search committee.

It is hard to say, based on these general considerations, what the optimal level of participation in the pool might be. The number has varied over time; when the pool began to operate in 1972, it had only five members, but its size has grown steadily over time, and at present all Justices except one participate. Whatever the optimal membership, it seems likely that zero non-participants is too little, and the Court is now uncomfortably close to that extreme. The larger point is that understanding the mechanism of interdependent search plus independent evaluation at least identifies the variables that determine the optimal setup of the pool.

Legislative committees

Legislatures are not often thought of as epistemic decision-making institutions. Rather, legislatures often act as a kind of political marketplace for bargaining between the major political parties, who have different fundamental preferences over major policies. Yet within parties, it is entirely plausible that a great deal of epistemic decision making takes place. Although parties are themselves coalitions, especially under first-past-the-post voting systems, legislative co-partisans are much more likely to share fundamental preferences with one another than with members on the other side of the aisle. The co-partisans have common aims – perhaps to promote the public good, perhaps to stick the other party in the eye – and their problem is to aggregate differing information and beliefs so as to achieve their common aims.

Enter legislative committees. There are many different theories of committees, all of which seem to capture some truth; however, we will focus on the implications of one such theory, the partisan control account of Cox and McCubbins (2007). On this account, political parties control committees. In contrast to interest-group accounts, which picture legislators self-selecting onto committees, the partisan control view holds that party leaders select committee members to promote partisan interests. In contrast to informational accounts, which see committees as serving the interests of the median member of the whole legislature (who has decisive power under simple majority rule), committees serve the interest of the median member of the majority party, who has decisive power in selecting the internal legislative leaders, who in turn control the composition of committees.

It is at least compatible with this view to suppose that the partisan majorities who select and direct committees face an epistemic problem, conditional on the common preferences of co-partisans. Committee members search for policies that will promote the majority party’s preferences, perhaps because the partisan majority believes those policies best for the nation, perhaps because they are politically constrained by constituents to adopt certain policies or block others, or perhaps to embarrass the other party or split the other party’s internal coalition. Whatever the motivation, the task is epistemic in that a partisan majority has (much of the time, on many issues) common preferences but dispersed information and differing beliefs about how to satisfy those preferences.

How can the party leaders identify and agree upon policies that will attain their ends? We suggest that the partisans might do well to imitate the bees, and in some respects already do so. They might set up a subgroup of the party membership to serve as a search committee. Individual members of this search committee would, in effect, roam the policy space to find politically useful legislative proposals. The members would then advertise any proposals identified as potentially useful, and would attract support from other co-partisans to the extent that, after inspecting the advertised candidates, they believe a given proposal is indeed politically valuable. If support reaches some critical threshold in the search committee, the committee would report out a bill embodying the proposed policy and, if politically feasible, the majority party would enact the bill into law.

As described, this process displays the combination of interdependence and independence characteristic of the bees’ decision making. The co-partisans act interdependently at the agenda-setting stage in which useful candidate proposals are identified by individual members of the search committee. Committee members are more likely to give serious consideration to candidate policies identified by other committee members, rather than searching the policy space in a strictly individual fashion, without regard to the recommendations of others. Conditional on investigating alternatives proposed by other members, however, the members exercise independent judgment about the quality of those alternatives. The ultimate selection among the candidate policies is determined by independent assessment of alternatives generated interdependently.

Real-world legislatures are not so different from this model process. We can understand legislative committees as searching the policy space for politically advantageous proposals, and then exercising conditionally independent judgment on a set of alternatives generated in an interdependent fashion. Prescriptively, from the standpoint of the majority party, the key question about this decision-making mechanism is whether the costs of interdependence at the agenda-setting stage exceed the benefits. Recall that the main cost is the possibility of premature herding towards a bad or at least suboptimal alternative because there has been insufficient exploration of alternatives. The main benefit is that insufficient levels of interdependence at the agenda-setting stage tend to produce a failure of consensus. Where the opportunity costs of inaction are high, as in the bees’ environment, avoiding this failure of consensus becomes a collective imperative.

For party leaders, an important implication is that interdependence at the committee agenda-setting stage becomes more valuable as the costs of inaction increase. Imagine a political environment in which the majority party will suffer, politically, if it is perceived as running a “do-nothing Congress.” At the early stages of the n-year legislative cycle, party leaders and committee chairs will do well to afford individual committee members more freedom to search out proposals that will put the opposing party in an awkward position, or will promote the majority’s platform. This freedom will slow down the process of consensus formation, but result in an increase in the expected quality of the eventual consensus, from the majority party’s point of view. As the election cycle nears its close, the opportunity costs of inaction increase, because the costs of being charged with running a do-nothing Congress increase. Party leaders should tighten up the process of search and resolution by requiring greater interdependence among committee members. Committee consideration should focus on the alternatives already identified, although committee members should be allowed to exercise independent judgment among those alternatives.

The comparison should not be pressed too far. How much are rank-and-file legislators really like scout bees? And do Senators bear anything more than a superficial resemblance to drones? A key difference between the insect hive and Capitol Hill is that legislatures are more specialized internally. Rather than having a single search committee to make a highly consequential collective decision (such as the choice of a new nest site), legislatures have multiple search committees each assigned to a different area within the total policy space. Moreover, these committees have partially overlapping jurisdictions, and the boundaries between their jurisdictions may be fuzzy.

Yet this jurisdictional fuzziness may work well, as judged against the bees’ mechanism. As in the certiorari pool, where the optimal level of participation is not 100%, so too it is not desirable, from the standpoint of party leaders, that there be complete interdependence at the committee agenda-setting stage. Complete interdependence would eliminate all noise from the system, and thus eliminate any prospect for members to stumble upon new and highly advantageous proposals not found by others. Jurisdictional fuzziness can introduce some desirable noise by making it possible, although unlikely, that a member from another committee will stray into the policy space and stumble upon a valuable proposal that members of the principal committee have overlooked. Of course, we do not suggest that the jurisdictional overlap and fuzziness of typical legislative committee-structures is best explained on these grounds; it arises for a number of political and historical reasons, rather than on the basis of any considerations of optimal collective decision making. Yet in light of the bees’ arrangement, jurisdictional overlap and fuzziness may be associated with some epistemic benefits.

Searching for leaders: Executive appointments in firms

The bee’s search-and-decision mechanism is most advantageous to the group when the decision-making environment combines high stakes (making independence at the voting stage valuable) with high opportunity costs of indecision (making interdependence at the agenda-setting stage valuable), while the agenda is relatively open-ended. Plausibly, the search for leaders in for-profit firms and non-profit organizations presents just such a decision-making environment. First of all, the set of possible candidates is not so easy to identify. The stakes of the choice are high, because leadership and charisma are scarce resources whose presence or absence can make or break institutions, and because firms and organizations tend to search for new leadership in periods of crisis, in which routine decision making is not viable and executive decisions are particularly consequential. The opportunity costs of failing to reach consensus on the selection of a new leader are also high under such conditions, because passivity and inaction are often the worst possible strategies for institutions in crises; it is better to have a strong hand at the helm than to drift in treacherous waters, even if it is unclear which way it is best to go.

The implication is that leadership searches in firms and organizations, especially in crisis conditions, should attempt to combine interdependence and independence in roughly the ways we have outlined. We will focus on the structure and procedures of executive search committees in universities. When universities select new leadership – say, the university’s President – the process typically involves a search committee. At Harvard University, for example, the search committee that selected Drew Gilpin Faust as President consisted of the six members of the University Corporation, plus three members of the University’s Board of Overseers. University search committees often operate in a secretive fashion, so it is difficult to know how they make decisions. However, the pool of suitable candidates is usually not so easily defined, and in the environment in which Faust was selected – in early 2007, as the financial crisis became ever more severe – the stakes were high and the opportunity costs of deadlock serious.

In such an environment, leadership search committees should engage in interdependent agenda-setting and independent assessment. Interdependent agenda-setting will mean that individual members of search committees tend to focus their attention on candidates previously proposed by other members of the search committee, rather than roaming the space of candidates purely on their own. The committee members who propose candidates early in the process may have predominant influence in setting the agenda. Yet this does not entail interdependence at the stage of evaluation. To the contrary, committee members should decide with strict independence whether the candidates on the agenda meet the threshold set by the group’s established criteria. Procedural mechanisms such as the use of secret ballots within the committee can maximize independence under certain conditions.

In this picture, there are twin evils to be avoided. On the one hand, insufficient interdependence would result in an excessively protracted search process, as committee members would spend too much time searching for new candidates to put on the agenda, and devote too little attention to evaluating the candidates others have put forward. On the other hand, insufficient independence would produce informational cascades that might settle on a bad candidate. The optimum is a process that allows agenda-setters to structure the pool of candidates, yet subjects those candidates to fully independent evaluation – maximizing the chances of settling on a good candidate, with reasonable expedition.

Similar lessons apply to committee decisions on the award of symbolically important prizes, such as Nobel Prizes and other national or international recognitions of merit. In such decisions, the pool of candidates tends to be open-ended; the stakes are high due to the cultural, intellectual, political, or sometimes commercial repercussions of an award, and not meeting the deadline for an award is normally not an option. Here, too, committees are often secretive about their procedures, but our discussion suggests that – at least in cases of common fundamental preferences – they would do well by balancing interdependence in arriving at a list of nominated candidates with an independent assessment of these candidates’ merits.

Research and development in firms

Just as the honeybees’ mechanism can be applied to identify candidates for certain positions or awards, so it can also be applied to identify projects worth developing. Imagine a firm with a large staff of research experts – scientists, engineers, or other knowledge workers – who engage in two sorts of tasks, dividing time between them in some proportion. One task is to independently search the space of technically feasible innovations for potentially profitable innovations, and then to promote them to other researchers. Another task is to assess potentially profitable innovations promoted by colleagues. In the second task, assessment is independent, but the choice of innovations to be assessed is interdependent. By promoting an innovation, researchers set the epistemic agenda for colleagues, who suspend their independent search and decide whether the proposal meets some threshold of plausibility. Proposals that attain sufficient support in the research group are kicked upstairs for further assessment by higher management.

How should this process be structured to maximize the firm’s expected utility? From the standpoint of the firm, the optimal allocation of time by each individual researcher will not be either of the corner solutions – either the one in which each researcher spends all her time independently searching for profitable innovations, or the one in which each researcher spends all her time assessing innovations proposed by others. In the latter case, there will be no innovations to assess – not everyone can be reactive, or there will be nothing to react to – while in the former case, no consensus will form and the firm’s collective resources will be scattered too widely across different projects. The optimal time allocation balances independence and interdependence.

There are several examples of highly creative firms that use an optimizing mechanism of this sort. In the 1950s, 3M allowed its research staff to devote 15% of their time to independent projects, whose results would belong to the firm. Famous innovations resulted, including Post-It Notes and masking tape. More recently, Google has a similar policy at the 20% level, which has been credited with producing Gmail and Google News. ¹³

This picture is a heroic simplification, because for-profit firms are often cited as examples of groups of actors with conflicting fundamental preferences. Standard principal-agent models of such firms begin with the premise that the lower-level agents have preferences that diverge from the preferences of the principal – the firm’s leadership, somehow defined – so that researchers may want to slack off, or to research questions that are of maximal interest to them rather than of maximal expected utility to the firm, and so on. However, the picture we advance has some utility to the extent that compensation mechanisms, incentive schemes, informal norms, or selection and screening at the hiring stage align the interests of researchers with the interests of the firm and its principal(s). Where that is so, it is not impossible to understand the optimal research and development process in for-profit firms along the lines we suggest, and to reconfigure the research and development process within actual firms accordingly.

Basic research

To illustrate the limits of the analysis, we suggest that there is a large domain in which the social utility of the bees’ decision-making mechanism is more limited: basic research, especially in the natural sciences. By basic research we mean research that has no currently foreseeable applications or direct payoff for applied sciences, such as engineering. Although society does well to fund a portfolio of basic research, some small fraction of which will pay off handsomely in the long run, there is no expectation that the payoff will materialize in the short run. Theoretical research in physics is the standard example.

Importantly, the process has both an epistemic dimension and a dimension of collective choice. The epistemic dimension is that we envisage basic research as a process of searching for theories about (some aspect of) the world, and we assume that some theories are objectively correct while others are objectively incorrect. The collective choice component can be understood either from the standpoint of funding institutions, or from the standpoint of scientists themselves. For funders, typically panels or other groups, the problem is to fund a portfolio of basic research that maximizes net present value to the funding institution or to society generally. For scientists, the problem is to generate a set of possible theories and then to reach consensus on the ones that are true, while collectively rejecting those that are false. As we will see, these twin collective aims – theory generation and theory sifting – trade off against one another.

In basic research, the bees’ mechanism would entail that researchers focus their attention primarily on questions identified by other researchers, and then proceed to address them independently. This has indeed been the tendency in basic research, which is increasingly conducted in teams across the natural sciences. Such teams in effect focus on questions identified by a leader or head, or on questions that have become fashionable, having received attention from other researchers; team members do not ask whether the question is the right one to pursue, only how, if at all, it can be answered. Although there are powerful institutional and individual incentives to form such teams, their social utility is an open question.

Our analysis suggests that an excessively high degree of interdependence at the agenda-setting stage of basic research is undesirable. The reason is that the opportunity cost of failing to reach consensus on basic theories is relatively low, where one assumes as we do that consensus on those theories will have no immediate payoff. Basic research is a long-run enterprise, in which it is better that things be settled right, eventually, than that things be settled today. The cost of failing to reach consensus – the main cost of low interdependence – is typically a lesser concern; the greater concern is, or should be, that excessive interdependence may leave some very promising theories sitting about undiscovered, because no one has been searching for them.

Our point is not that basic research teams are inaccurate, or unreliable. With the bees’ combination of high interdependence and high independence, basic research teams will make accurate assessments of theories put on their scientific agenda by team leaders, yet some excellent theories may go unconsidered by anyone. With lower interdependence, more scientists will roam the theoretical space alone or in smaller teams; more theories will be explored, but by fewer people in each case. That retards the generation of consensus on the theories considered, yet reduces the number of true theories overlooked altogether.

By the nature of the case, it is hard to know whether there are many excellent theories waiting to be found, and if so where they might be (if one knew those things, one would already be in position to find the theories). However, we can motivate our view by pointing to the domination of string theory within theoretical physics, as a possible example of the costs of interdependence. In the current generation, “it is virtually impossible (in the U.S.) for someone not working within [the string theory] paradigm to be hired as an assistant professor [of physics] at a major research university” (Elster, 2009: 19–20). According to critics, theoretical physics is focused to the point of obsession on evaluating and expanding one particular theory or family of theories while other approaches go unexplored. Meanwhile, a growing number of people have begun to question whether string theory is even a scientific enterprise at all, given the difficulty of using the theory to generate implications that are both testable and unique to the theory (Elster, 2009: 20; Woit, 2006). If the criticism is right, theoretical physics suffers from excessive interdependence of the research agenda. Similar social dynamics among researchers may be responsible for the excessive disciplinary rigidity of which mainstream economics is sometimes accused. ¹⁴

Statistical groups: Ratings and individual choices

We may also consider an extension, from the judgments made by actual groups to the virtual judgments made by statistical or notional groups, and to individual choices that rely upon the judgments of those statistical groups. An example of what we have in mind is the website “Rate My Professors,” ¹⁵ where students can see, for any given professor, an “overall quality” score that averages all ratings, as well as composite scores for “easiness” and “hot(ness).” Students who use websites of this sort to decide where to allocate their course time are implicitly relying upon the collective judgment of a statistical group. Importantly, not all professors have ratings. Which professors will be rated depends on the decentralized choices of other students, who in effect advertise the (high or low) quality of a professor by choosing to rate them. That choice in turn influences the choices of later students to take or not to take a given course; the students who do so may then record an independent assessment of the professor’s quality, which will in turn influence the choices of yet later students, and so on. There is interdependence among students at the stage of deciding which course to take, yet independence (ideally, at least) at the stage of rating.

More generally, a similar combination of interdependence and independence can appear in a broad range of individual choices influenced by the decisions of earlier participants to advertise (including in negative terms) the quality of the choices. The relevant category here is the rating system, which can be distinguished from a ranking system; the former is decentralized while the latter is inherently centralized. Under a ranking system, some individual or group attempts to reach a synoptic overview of the relevant choices – an exogenously defined set – and compares them all with one another in order to arrive at an overall ranking. Under a rating system, by contrast, collective judgments emerge from the decentralized action of participants who search the space of possible choices, advertise for or against candidates they like or dislike, and thereby influence the choices of other participants whether or not to consider or sample the relevant goods. Although individuals decide which options to consider or pursue by considering the aggregate judgment of a virtual group, their assessment of those options is (ideally) independent. In this respect, the growing prevalence of rating systems for consumer goods – particularly experience goods such as films, ¹⁶ hotels, vacations, ¹⁷ and gourmet restaurants ¹⁸ – shows that homo sapiens has already begun to do as the hive does, although many millennia after the hive perfected its mechanism of decentralized search.

Interdependence, independence, and institutional design: Veiling mechanisms

In any decisional environment that satisfies the three conditions we have described – an open-ended agenda, high stakes, and high opportunity costs of indecision – the challenge for institutional designers is to balance interdependence and independence. The former is beneficial at the stage of agenda-setting, in order to coordinate on suitable agenda items, while the latter is beneficial at the stage of evaluation of those items, in order to maximize the accuracy of the resulting decision.

A difficulty, however, is that interdependent agenda-setting might spill over into the stage of evaluation. Suppose, for example, that in an executive search committee one committee member is known to have placed a candidate on the agenda because he or she considers the candidate to be of very high quality. That knowledge might then affect the others’ substantive evaluations of the candidate, reducing their independence. Indeed, taking into account the information embodied in the agenda proposals of others may be individually rational for any given participant, so long as the benefits of incorporating the information are greater than the additional cognitive costs. ¹⁹

There is, thus, a potential disanalogy between bees and humans. In the bees’ mechanism, individuals do not directly observe others’ full quality assessment, which is encoded in the duration of their dances. Dance duration indirectly affects the probability that other bees will investigate a given site, but there is no direct spillover of information from the agenda-setting stage to the evaluation stage. Observing duration may exceed the cognitive capacities of any individual bee. Thus the bees are indirectly advantaged by their low cognitive capacities, combined with the structure of their decision-making mechanism. Paradoxically, humans’ superior cognitive capacity enables them to infer information from agenda proposals, compromising independence; humans thus face the spillover problem.

The proper response to the problem, however, is not to declare the bees’ mechanism irrelevant to human collective choice in epistemic contexts. Rather, the challenge is to replicate the conditions for the bees’ success by employing strategies of mechanism design. The same human ingenuity that creates the spillover problem can also ameliorate it, by enabling human institutional designers to develop mechanisms that shield the stage of independent evaluation – in whole or at least in part – from the information required for interdependent agenda-setting.

The precise mechanisms that can enforce such shielding are highly context-specific. However, we will describe one general class of veiling mechanisms ²⁰ that in one way or another limit the information held by participants when evaluating proposals others have placed on the agenda. Such mechanisms will create a barrier between interdependence and independence to prevent spillovers that might compromise the latter. By depriving human decision makers of information they might use – quite rationally – in ways that compromise independence, veiling mechanisms indirectly replicate the bees’ lower cognitive capacities, with the paradoxical result of improving group performance.

Here are some examples. In executive search committees, one might establish an impartial officer to serve as a depository for proposals to place a candidate on the agenda. The officer will inform members of the bare fact that some other member has proposed consideration of the candidate, but nothing more. Whereas under ordinary procedures, each member of the committee will know who proposed the candidate and will hear a formal or informal presentation of reasons in the candidate’s favor – information that might compromise independence at the stage of evaluation – the laundering of proposals through an intermediary acts as a partial veil that reduces the flow of compromising information.

Likewise, the Supreme Court might experiment with veiling procedures. Under current practice, the pool memorandum that recommends a grant of certiorari often contains a detailed statement of the case and of the parties’ arguments, and lists the name of the authoring law clerk. Although we doubt that this information has a large compromising effect on the independence of evaluation by Justices and other clerks, especially since cases are ultimately decided only after full briefing and argument, one might eliminate the law clerks’ names and at least some of the other information to minimize the risk of spillover. In these and other settings, spillover is not only a problem, but also an opportunity to design mechanisms that replicate and exploit the advantages of the bees’ procedure.

Epistemic agenda-setting

Machiavelli ([1513]1996) observed that “a multitude without a head is useless.” In his motivating example, a group of plebeians who threatened to secede from Rome proved entirely incapable of negotiating with the patricians, because the plebeians had no leader to make proposals to them and speak for them. Although Machiavelli did not clearly distinguish between the aggregation of judgments and preferences, his observation holds in either setting, insofar as he is pointing out that it is often costly for decision-making groups to structure their own agendas. Even when a decision-making group shares all fundamental preferences in common, it must make choices between alternatives, and the alternatives must come from somewhere. Given realistic constraints on the time and cognitive capacities of the group, it is not feasible for all members of the group to put as many proposals as they see fit on the agenda, and then for the group to vote on all proposals. Instead, some member or members must act as epistemic agenda-setters who narrow the range of options.

From another standpoint, however, Machiavelli’s claim is misleading. The epistemic agenda-setter on a given issue need not be a “head” or leader, in the sense of an individual member, specified ex ante, who possesses agenda-setting authority across the board. Rather the agenda-setter can be a rank-and-file member, or the member of a search committee, who advertises the quality of a given option to other members, who in turn coalesce around a particular option and influence the choice of the group as a whole. In the honeybees’ procedure, there are epistemic agenda-setters, but they are otherwise unremarkable rank-and-file bees. It is pragmatically necessary that there be epistemic agenda-setters, but that is a different topic than leadership.

Many issues lurk here, and there is much to be explored. The honeybees’ procedure merely illustrates that work on collective epistemic decision making, in economics and rational choice theory more generally, has neglected the issue of epistemic agenda-setting, and has generally rested content with models that treat the options for decision as exogenous. The next generation of epistemic models should relax this assumption.

Time, truth-tracking, and collective search

A second and related issue is that Condorcetian models of collective epistemic decision making are excessively static. Those models illuminate the idea that collective decisions can produce correct or incorrect judgments, relative to the common aims of the group; in that modest sense, the models attempt to show conditions under which group decisions can “track the truth” (Estlund, 1993; Grofman et al., 1983; List and Goodin, 2001). Yet the Condorcetian models show little appreciation of the brute fact that searching for the truth takes time. Sensible groups will trade off the benefits of obtaining the very best answer against the opportunity costs of information exchange, deliberation, and possible failure to reach consensus, resulting in no group decision at all (in effect a decision for the status quo, which may be untenable or the worst option of all). By contrast, models of search processes in economics, behavioral economics and biology are acutely sensitive to the opportunity costs of search and to tradeoffs between speed and accuracy. Yet those models frequently involve a single decision maker, perhaps a consumer, and thus abstract away from the crucial collective epistemic problem: multiple decision makers have different beliefs and information, which must somehow be aggregated through an optimizing procedure.

The honeybees’ decision procedure lies at the intersection of these problems. The honeybees face a problem of collective search in an epistemic context. Some options are much better than others, given the common fundamental preferences of the group, yet different members of the group have different information than others, and the group’s problem is to pick an optimal strategy for collective search. Of course, the honeybees do not pick a strategy, either individually or collectively; yet natural selection has produced individual-level behaviors that cause the bees to behave in a collectively optimal fashion. For human purposes, the bees’ as-if collective search strategy is worth considering in any decision-making environment that resembles the bees’ environment. As we have tried to show, humans face similar environments in a range of institutional settings.

Comparative institutional analysis across species

The last point explains the sense in which, and the conditions under which, the hive provides useful lessons for humans. Precisely because individual bees have such low cognitive capacities, the apparent efficiency of their collective search strategy, in their environment, is all the more striking. To the extent that their aggregation mechanism exhibits efficiency, the human problem is to describe the features of their environment in suitably abstract terms – the terms of economic theory and decision theory – and then to ask whether there are similar human environments or decision problems to which the bees’ mechanism might be carried over. Nothing in this process of analysis, abstraction, and transposition requires drawing dubious analogies between humanity and the hive. Drawing lessons from the hive is merely a form of comparative institutional analysis across rather than within species. Just as institutional designers may observe other human organizations in the social world, to find institutional forms that would otherwise never have occurred to them, so too designers may observe the products of natural selection to broaden their repertoires.