Abstract
Majority-voting and the Condorcet Jury Theorem pervade thinking about collective decision-making. Thus, it is typically assumed that majority-voting is the best possible decision mechanism, and that scenarios exist where individually-weak decision-makers should not pool information. Condorcet and its applications implicitly assume that only one kind of error can be made, yet signal detection theory shows two kinds of errors exist, ‘false positives’ and ‘false negatives’. We apply signal detection theory to collective decision-making to show that majority voting is frequently sub-optimal, and can be optimally replaced by quorum decision-making. While quorums have been proposed to resolve within-group conflicts, or manage speed-accuracy trade-offs, our analysis applies to groups with aligned interests undertaking single-shot decisions. Our results help explain the ubiquity of quorum decision-making in nature, relate the use of sub- and super-majority quorums to decision ecology, and may inform the design of artificial decision-making systems.
Impact Statement Theory typically assumes that majority voting is optimal; this is incorrect – majority voting is typically sub-optimal, and should be replaced by sub-majority or super-majority quorum voting. This helps explain the prevalence of quorum-sensing in even the simplest collective systems, such as bacterial communities.
Effective decision-making is essential in all areas of human society and, more generally, for all organisms. A fundamental question in this context is when a group of decision-makers is superior to individual decision-makers and vice-versa [1-6]. Both in human and animal collective decision making, the Condorcet Jury Theorem is one of the key principles guiding our thinking about this question [7-14]. In a nutshell, for pairwise decision problems (e.g. disease-, lie- or predator detection), Condorcet’s Jury Theorem states that a group of decision-makers employing the majority rule is superior to individual decision-makers in contexts where individuals are relatively accurate (i.e. accuracy > 50%); conversely, individual decision-makers are superior to majority voting groups in contexts where individuals are relatively inaccurate (i.e. accuracy < 50%). Consequently, across a diverse range of fields ranging from organismal behaviour and human psychology to political sciences, two heuristics are commonly employed [7-12]: (i) groups of decision-makers outperform individuals only in contexts where individuals are relatively accurate and (ii) the majority rule is a powerful mechanism to reap the benefits of collective decision making. We here show that both statements are not true, and in doing so explain the ubiquity of quorum decision rules in the collective behaviour of humans and other organisms [15-21].
Over the past few decades, substantial research effort has focussed on two key explicit assumptions underlying Condorcet’s Jury Theorem, independence (i.e. judgments/votes by different members of the group are assumed to be independent from each other) and homogeneity (i.e. all decision-makers within a group are assumed to be identical, both in competence and in goals) [12, 16, 22-25]. We here focus on a third, implicit, assumption of Condorcet’s Jury Theorem, namely that decision-makers make only one type of error. This assumption stands in contradiction to the well-known fact that, when confronted with a pairwise decision problem like the one studied in Condorcet’s Jury Theorem, two different types of error are possible (i.e. false positive and false negative). Surprisingly, up to now, this basic and well-known feature of pairwise decision problems has not been taken into account when investigating Condorcet’s Jury Theorem.
We proceed as follows. We start by providing a brief summary of the basic model considered in Condorcet’s Jury Theorem. We then introduce an extended model that takes into account the fact that decision-makers can make two types of errors. Based on this extended model, we then show that Condorcet’s Jury Theorem makes several important predictive errors, which apply to the majority of decision scenarios. Moreover, majority voting is frequently suboptimal, whereas quorum-based voting with an appropriate quorum is always optimal. Thus, the simple majority threshold is only a special case of the more general quorum decision mechanism, in which optimal sub-majority or super-majority quorums are the rule rather than the exception.
CONDORCET’S JURY THEOREM: THE BASIC MODEL
Condorcet’s Jury Theorem considers a binary (pairwise) choice situation, in which decision-maker can choose between two actions, labelled +1 and –1. Each decision-maker is characterised by a single parameter a, corresponding to the probability of making a correct decision, or decision accuracy. Importantly, the decision accuracy a of each decision-maker is assumed to be conditionally-independent of the realised decisions of all other decision-makers.
Condorcet’s Jury Theorem now considers a group of identical decision-makers of size N that performs a majority vote. A simple combinatorial argument shows that if the accuracy of decision-makers is above 50% (i.e. a > 0.5), then the probability of making a correct choice increases with increasing group size and asymptotically approaches 1 [22]. Conversely, if the accuracy of individual decision-makers is below 50% (i.e. a < 0.5), then the probability of making a correct choice decreases with increasing group size and asymptotically approaches 0 [10]. This is because, as group size increases, the probability of the more probable decision (+ or -) also being the majority decision rapidly increases towards one.
These results have led to three key interpretations: first, pooling independent judgements is beneficial (i.e. improves decision accuracy) whenever individuals are relatively good decision-makers (a > 0.5). Second, pooling judgements is detrimental (i.e. decreases decision accuracy) whenever individuals are poor decision-makers (a < 0.5). Third, the majority rule is the appropriate mechanism to reap the benefits of collective decision making.
In the following, we show that each of these interpretations is incorrect. In particular, we show that (i) in cases where decision-makers are good (a > 0.5) majority decisions may decrease decision accuracy, (ii) in case where decision-makers are poor (a < 0.5) majority decisions may increase decision accuracy, and (iii) pooling independent decisions is always beneficial as long as an appropriate quorum-based decision rule is used. Taken together, these results show that, for a large proportion of decision scenarios, the simple majority decision rule performs poorly and gives incorrect predictions about group decision accuracy.
CONDORCET’S JURY THEOREM: AN EXTENDED MODEL
Condorcet’s Jury Theorem implicitly assumes that decision-makers can make only one error, that is, the probability of making an incorrect decision is 1 – a. This is in contrast to the well-known fact that – in pairwise decision problems – decision-makers can make two type of errors, false positives and false negatives [26, 27]. For example, an animal under predation risk may run away in the absence of a predator (false positive) or it may not run away in the presence of a predator (false negative) [28]. Similarly, a doctor screening for a disease may diagnose in the absence of a disease (false positive) or not diagnose in the presence of a disease (false negative) [29]. As we discuss below, the implicit assumption of Condorcet’s Jury Theorem is equivalent to assuming that decision-makers have an identical probability of committing the two errors – this is an important assumption that does not reflect the vast majority of real world decisions.
We here consider an extension of the above-described model, which takes into account the fact that decision-makers can make two types of errors. Again, decision-makers can choose between two actions, +1 and –1. Unlike in the basic model above, however, and consistent with standard decision theory for pairwise decisions, we now assume that the world can be in two states, state + and state –, corresponding to, for example, the presence and absence of a predator or the presence and the absence of a disease; state + holds with probability p and state – thus holds with probability 1 – p. Action +1 is the better choice in state + (i.e. it achieves a higher payoff), while action –1 is the better choice in state –; for example, running away is better than staying in the presence of a predator, while staying is better than running away in the absence of a predator. Consequently, and in contrast to the basic Condorcet model above, each decision-maker is now characterised by two parameters a+ and a–, corresponding to the probabilities of making correct decisions in state + and state –, respectively; this inevitably implies that individuals make two types of errors. As in the simple model above, where decision-makers have equal accuracies a and are independent, for any individual in a group of N individuals the probability of making a correct decision based on the true state of the world is conditionally-independent of the probabilities of other group members making the correct decision, given that same state of the world.
To relate our analysis to predictions made by applying Condorcet, we must define expected individual accuracy a, as used in Condorcet, in terms of our state probability p and state-wise accuracy parameters a+ and a–. The expected individual accuracy is thus
From equation 1 we can see that Condorcet implicitly assumes that accuracies in the two states of the world are equal since then p disappears from the equation; as shown in Supplementary Information, this occurs when both states of the world are equally likely, and the costs and benefits of classifications in the two states of the world are equal, although asymmetric decision problems can also result in equal accuracies (as can be confirmed with reference to Supplementary Information for Figure 1). Later, we explain how signal detection theory determines optimal values of a+ and a– for a given decision scenario where these assumptions are violated.
RESULTS
We are now ready to formalise when Condorcet gives incorrect predictions, and when simple majority voting is suboptimal. We begin by determining optimal values of a+ and a–. Given the prior probability p, the cost and benefits associated with the two states of the world, and the state-dependent characteristics of the cue(s) decision-makers base their decisions on, the optimal realised individual accuracies a+ and a– are derived by solving a signal detection problem (Figure 1). Signal detection theory has been applied repeatedly to hypothetical and real world decision problems, for example predator detection by foraging animals [28, 30]; diagnostic decision making [4, 31]; and lie detection [32]. As described in the appendix, for simplicity and tractability our analysis is conducted for the simplest signal detection problem, determining which of two distributions a single scalar random variable is drawn from; this could, in the example of a predator detection problem, be the instantaneous volume of a sound heard by a forager. In this case signal detection theory enables us to find a Receiver Operating Characteristic (ROC) curve of optimal a+, a– value pairs, dependent on decision difficulty (Figure 1; [26, 33]); the optimal point of this curve gives us a unique pair of a+ and a– values dependent on costs and benefits of different decision outcomes, and the prior probability of the + state, p.
We next make the observation that, to simultaneously improve both true positive and false positive rates, a group must choose a quorum q that lies between these two values [29]; that is, an optimal group must choose q such that
The intuition behind this result is that group accuracy will converge on the appropriate accuracy for the true state of the world, as group size increases; thus, by setting a quorum between these two accuracies the true state of the world can be determined with high probability for sufficiently large groups. For further details see [29]. Since the simple majority quorum q = 1/2, assumed by Condorcet, only satisfies this inequality when both a+ > 1/2 and a– > 1/2, when either of these conditions are violated then both simple majority decisions and Condorcet-based reasoning will be deficient. Furthermore, we should never see true positive rate (a+) less than false positive rate (1 - a–) (dark grey region of ROC space in Figure 1), since such a decision-maker could simultaneously improve both their true and false positive rates simply by inverting their decisions and moving above the diagonal. Therefore, the ROC space is divided into two meaningful regions: in the first a+ > 1/2 and a < 1/2 (white region in Figure 1), so simple majority voting is asymptotically-optimal as group size increases, and Condorcet-based predictions are correct. In the second, a+ < 1/2 and a– > 1/2, or a+ > 1/2 and a– < 1/2, while ensuring a+ > 1 - a– (light grey regions in Figure 1); in these regions simple majority decisions will be sub-optimal, and Condorcet-based reasoning will be erroneous.
We now describe the systematic errors Condorcet leads to when faced with these decision scenarios. These errors can also be described in terms of false positive (frequently referred to as type I) and false negative (type II) errors, in predicting the performance of the Condorcet majority rule, hence we label the errors in such terms.
Error Ia: Condorcet predicts group accuracy approaches 1, but majority groups do not
This error occurs when majority voting is sub-optimal (a+ < 1/2 and a–> 1/2, or a+ > 1/2 and a– < 1/2; light grey regions in Figure 1, as described above), and when expected individual accuracy a > 1/2, since under this last condition individuals on average make more correct decisions than incorrect decisions, and Condorcet thus predicts that group accuracy approaches 1 as group size increases [10, 22]. From equation 1 this requirement is thus that
While Condorcet predicts group accuracy (denoted ā) approaches 1, that is in fact since the majority quorum q = 1/2 is either below 1 – a–, or above a+, contra inequality 2; thus the group converges to making completely correct choices in one state of the world, and completely incorrect choices in the other state of the world [29]. Hence, group accuracy converges to
The wide range of decision scenarios in which Condorcet makes this predictive error are illustrated in Figure 2a. An example of this error is presented in Figure 3a; Figure 3b illustrates how the error can be avoided by choosing an appropriate quorum, in this case a super-majority quorum.
Error Ib: Condorcet predicts group accuracy approaches 1, but majority groups are worse than individuals
In error Ia, while group accuracy does not converge to 1 with increasing group size, nothing is said about whether or not groups are better than individuals. Error Ib refines error Ia, by showing that there are cases where Condorcet predicts group accuracy approaching 1, but groups actually have lower accuracy than individuals. These cases are found by refining the conditions given in error Ia to include the additional condition that the group accuracy converged to (equation 5 above) is less than individual expected accuracy (equation 1). This gives the conditions and
The wide range of decision scenarios in which Condorcet makes this predictive error are illustrated in Figure 2b. An example of this error is presented in Figure 3c; Figure 3d illustrates how the error can be avoided by choosing an appropriate quorum, in this case a super-majority quorum.
Error II: Condorcet predicts group accuracy approaches 0, but majority groups do not
In error II individual expected accuracy is below 1/2, thus Condorcet predicts that group accuracy should converge to 0; while groups using the majority decision rule do decrease in accuracy, they converge to a non-zero group accuracy given by equation 5 above. To find cases where this occurs we simply solve for when individual expected accuracy a < 1/2. From equation 1 this gives us the conditions and
Decision scenarios in which Condorcet makes this predictive error are illustrated in Figure 2c. An example of this error is presented in Figure 3e; Figure 3f illustrates how the error can be avoided by choosing an appropriate quorum, in this case a sub-majority quorum.
Note that it is not possible to find an ‘Error IIb’ case that parallels Error Ib; that is if Condorcet predicts that group accuracy approaches 0, majority groups will always be worse than individuals and never better, even if their group accuracy remains positive. This is because it is not possible simultaneously to satisfy the conditions just given (inequalities 8 and 9), and the opposite of the conditions (inequalities 6 and 7) given in Error Ib (i.e. the conditions that group accuracy converged to is greater than individual expected accuracy), as can be confirmed by algebra.
Majority voting is usually suboptimal
Combining the cases in which Condorcet makes one of the above described predictive errors, Figure 2d illustrates when Condorcet will make at least one error in predicting the performance of decision-making groups using majority decision-making. This set also corresponds to the set of decision scenarios in which majority decision-making is sub-optimal, and is outperformed by an appropriately set sub- or super-majority quorum. Figure 2d shows that Condorcet is optimal in far fewer decision scenarios than those in which it is outperformed by an appropriate quorum rule. Furthermore, our analysis also relates decision-ecology to the requirement for sub- or super-majority quorums. From inequality (2), sub-majority quorums are required whenever a+ < 1/2, which corresponds to the upper-left area of Figure 2d; in contrast, super-majority quorums are required whenever 1 – a+ > 1/2, which corresponds with the lower-right area of Figure 2d (see Supplementary Information). Thus, whenever the positive state of the world + is rarer, and / or false positives are relatively expensive compared to false negatives, then a sub-majority quorum should typically be employed, while the converse holds for super-majority quorums.
DISCUSSION
We have shown that simple majority-based collective decisions are often suboptimal, and that consequently sub- or super-majority quorums should frequently be employed by groups of ‘like-minded’ individuals combining independent decisions. Our results are important for two reasons. First, the majority-based decision rule, and Condorcet-based reasoning, has been widespread in several major branches of collective decision theory. In the animal behaviour community the Condorcet prescription on individual accuracy exceeding ½ has been used to recommend when opinions should be pooled [10], and when experts should be favoured over group opinions [34]. Other authors have also invoked Condorcet and majority voting as the gold-standard for collective decision-making [7, 12, 35, 36]. All such studies implicitly neglect the reality that most decisions have two types of errors.
Second, while quorums have been widely studied in collective decision theory, we here present a comprehensive theory that may help explain their prevalence. Sub- and super-majority quorums have been considered theoretically [16]; Sumpter and Pratt implicitly assume only one type of error need be considered, and proceed from that point with their analysis, referring to quorum functions such as managing speed-accuracy trade-offs [37]. Conradt and Roper [38] in their review refer to true and false positives, but in explaining quorum usage refer back to earlier analysis as a mechanism to avoid extreme group decisions where individual fitness interests do not completely align [39]. Ward et al. [40] consider quorum use for facilitating information transfer in shoaling fish, yet ignore the possibility of different error types despite their great asymmetry in the scenario studied, predator detection. In contrast to these previous analyses here we have shown the fundamental importance of quorums in one of the simplest possible collective behaviour scenarios, single-shot collective decisions in homogenous groups, where individuals’ interests are aligned, and decision-making abilities do not differ. Thus, we might expect the use of quorums to be widespread in the natural world, even in the simplest of decisions. The widespread use of quorum sensing in bacteria provides evidence of this [41], and may prove a particularly good testbed for our theory given its binary nature and asymmetric state priors and error costs, although evolutionary conflicts of interests within bacterial communities may result in confounds. Moreover, as humans have been shown to employ quorum rules and adaptively adjust the associated quorum thresholds, human decision-making experiments may also provide a powerful approach to test our predictions [42, 43].
It is surprising that signal detection theory has seen relatively scant application to collective decision-making. Wolf et al. [29] noted the potential for different error types in identifying how quorums can improve collective decision accuracy; while motivated by signal detection theory they did not directly apply it to optimise the individual decision-makers and relate this back to group behaviour. Kirstein & Wangenheim [44] also noted the possibility for independent error types, again with reference to signal detection theory; they noted the potential for Condorcet to make the same qualitatively incorrect predictions that we note here, however they did not apply the relevant theory to delineate the situations under which Condorcet reasoning is incorrect, nor did they consider the possibility for quorum-based decision rules to rescue these situations. Sorkin et al. [45] applied signal detection theory to Condorcet-like models with varying super-majority quorums, but did not find the mechanism by which group decisions can be optimised [29]. Laan et al. [46] noted that Condorcet and majority voting can be suboptimal, and considered ways in which voting can be improved; while they discussed signal detection theory and voting mechanisms they neither explicitly considered error types, nor quorum thresholds, focussing mainly on correlations between decision-makers and the impact of the cost function used, as well as suggesting a data-drive machine learning approach to improving collective decision-making rules.
In contrast to earlier work, by applying signal detection theory we have simultaneously shown here both how fragile Condorcet and majority-voting are, and how the use of sub- or super-majority quorums should relate to decision ecology. Although simple collective behaviour models have been well studied and highly influential, our results, and others relaxing other assumptions of such models [25], indicate the subtlety that may be revealed in collective decision-making by a richer consideration of individual decision theory. Our work may additionally have the potential to inform work in the design of artificial decision-making systems machine learning and robotics; for example, in the field of ensemble learning, in which predictions from multiple weak classifiers such as neural networks are combined to improve decision accuracy, variable quorums, referred to as ‘threshold shift’, are used [e.g. [47]; however majority voting is still among the simplest and most ubiquitous vote fusion rules discussed [[48, 49]]. Hence, we suggest that our simple perspective on how to combine votes may also yield technological insight.
Appendix – Signal Detection Theory Reveals Cases where Condorcet Predictions are Incorrect
To determine if a group of optimal decision-makers could exist such that simple application of Condorcet would lead to erroneous predictions, but a quorum rule would allow optimal opinion pooling, we consider groups of identically-capable individual decision-makers (as assumed in Condorcet), modelled as making optimal decisions using signal detection theory in order to classify continuous signals arising from one of two possible normal signal distributions of equal variance [26]. That is, decision-makers are faced with a signal and must choose an optimal signal threshold, x, in order to classify signals as being drawn from either of the two possible normal distributions. Each distribution has a different mean (μ+ versus μ–) but the same standard deviation (H). In a natural setting the signal could represent information as to whether a predator is present or not, for example, with the two states of the world, predator present versus predator absent, having different distributions for this signal.
The optimal decision threshold x is chosen to minimise the expected loss for the decision-maker (or maximise the expected gain). The expected loss from a decision is where CTP, CFN, CTN and CFP are respectively the costs of true positives (correctly classifying state +), false negatives (incorrectly classifying state –), true negatives (correctly classifying state –) and false positives (incorrectly classifying state +). Thus, an optimal decision-maker should minimise equation A.2 by appropriately choosing the decision threshold x. Since, given x, and where Φ is the cumulative distribution function for the normal distribution, the optimal x can be found by substituting (A.3) and (A.4) into (A.2), differentiating the resulting equation and solving for zero [26]. Note that the optimal threshold x, and thus the optimal accuracies under state + and state – of the world, a+ and a– respectively, are affected both by class imbalance (p ≠ 1/2), and by asymmetric error costs (CTP - CFN ≠ C TN - C FP).
Acknowledgements
We thank Gavin Brown and Nikolaos Nikolaou for helpful discussions on ensemble learning theory. JARM was funded by the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement number 647704).