Add Journal to My Library
The Review of Economic Studies
, Volume Advance Article (3) – Dec 7, 2017

39 pages

/lp/ou_press/a-conversational-war-of-attrition-5MgBMJboLj

- Publisher
- Oxford University Press
- Copyright
- © The Author 2017. Published by Oxford University Press on behalf of The Review of Economic Studies Limited.
- ISSN
- 0034-6527
- eISSN
- 1467-937X
- D.O.I.
- 10.1093/restud/rdx073
- Publisher site
- See Article on Publisher Site

Abstract We explore costly deliberation by two differentially informed and possibly biased jurors: A hawk Lones and a dove Moritz alternately insist on a verdict until one concedes. Debate assumes one of two genres, depending on bias: A juror, say Lones, is intransigent if he wishes to prevail and reach a conviction for any type of Moritz next to concede. In contrast, Lones is ambivalent if he wants the strongest conceding types of Moritz to push for acquittal. Both jurors are ambivalent with small bias or high delay costs. As Lones grows more hawkish, he argues more forcefully for convictions, mitigating wrongful acquittals. If dovish Moritz is intransigent, then he softens (strategic substitutes), leading to more wrongful convictions. Ambivalent debate is new, and yields a novel dynamic benefit of increased polarization. For if Moritz is ambivalent, then he toughens (strategic complements), and so, surprisingly, a more hawkish Lones leads to fewer wrongful acquittals and convictions. So more polarized but balanced debate can improve communication, unlike in static cheap talk. We also show that patient and not too biased jurors vote against their posteriors near the end of the debate, optimally playing devil’s advocate. We shed light on the adversarial legal system, peremptory challenges, and cloture rules. 1. Introduction It’s not easy to raise my hand and send a boy off to die without talking about it first$$\ldots$$ We’re talking about somebody’s life here. We can’t decide in five minutes. Supposin’ we’re wrong. $$\hspace{1.8in}$$ Juror #8 (Henry Fonda), Twelve Angry Men Economics is not in the business of disputing tastes. And few topics pique an economist so much as seeing how preference diversity is resolved. We have in mind juries, tenure cases, FDA panels, Federal Open Market Committee meetings, etc. In each case, partially informed individuals share an imprimatur to dispassionately arrive at the truth. Two key features of the meetings are that (1) the search for truth falls short of certainty since the debate is a costly endeavour for all involved, and (2) the debaters might disagree on the costs of different mistaken decisions. This article develops a new model of debate in which both the length of debate and the wisdom of its decision reflect the debaters’ biases, delay costs, and quality of information. In the cases we envision, the Bayesian parable of information misses its complexity and intrinsic detail. Debaters cannot simply summarize their insights in one likelihood ratio, and not surprisingly, we do not see this happen. In the movie “Twelve Angry Men”, $$e.g.$$, each juror knew different aspects of the witness testimony; on an FDA panel, each member may specialize in different technical aspects of a proposed new drug. Intuitively, debaters each possess a myriad of “pieces of a puzzle”. Our debaters are also duty-bound to arrive faithfully at a decision, and so any argument must be verifiable; but this in practice forces everyone to explain their logic, and carefully adduce all facts: A federal juror must “solemnly swear” to ensure “a true deliverance $$\ldots$$ according to the evidence”. Naturally, it takes more time to explain it at a finer grain. In this story, even debaters with identical preferences take time to distill information to their peers. To capture all these features of debate, we explore a simple as if parable. In lieu of a complex signal space and boundedly rational debators, we substitute coarse communication in a standard Bayesian model, and assume that delay is explicitly costly. We specifically explore the dynamics of costly deliberation by two jurors who must agree upon a conviction or acquittal verdict. We assume that any open non-committal communication has already passed, and instead focus on the dispositive communication phase, where conversation “gets real”. In this voting parable, each juror incurs a delay cost any period a vote is cast. Two concurring votes (“moved” and “seconded”) irreversibly seal the verdict; otherwise, voting continues. We begin with two biased and partially informed jurors, Lones and Moritz. What emerges is an incomplete information war of attrition where the jurors alternatively argue for their natural verdict, conviction for the hawk Lones and acquittal for the dove Moritz, until one concedes. The coarse communication we study captures the spirit of our as-if parable, veiling the precise jurors’ types. For an equilibrium is a sequence of type intervals for each player (Theorem 1). A juror opposes his peer until his threshold surpasses his type, and then concedes. An equilibrium describes a zick-zack threshold path through the type space. Reminiscent of partial equilibrium analysis, triangular deadweight loss deviations from the diagonal measure the decision error costs—an error of impunity (wrongful acquittal) or miscarriage of justice (wrongful conviction). We then characterize all sequential equilibria in which jurors sincerely vote for their desired verdict. Sincere equilibria are indexed by their “drop-dead” dates (Theorems 2 and 3). In a deferential equilibrium, one player concedes by a finite date. Here, arguments end either by an equilibrium protocol or fixed cloture rule—such as in parliamentary debates. But our focal equilibrium has no certain last period. This communicative equilibrium intuitively corresponds to the finest grain parsing of the “complex signals” in our motivational story. It is the only stable equilibrium (Theorem 4) when jurors are equally patient and not too biased. Intuitively, deferring is not forwardly rational, since types who deviate can convey their powerful private signals. We next introduce a fundamental taxonomy of debating genres. When Moritz insists on his acquittal verdict, he incurs explicit delay costs. In a standard, private-value war of attrition, such costs are balanced by the strategic gains of outlasting his rival. Moritz’ incentives might well have this flavour: He might prefer to outlast every type of Lones planning to concede next period; for instance, this arises when Lones is very hawkish, and so pushes hard for conviction. Since these incentives are adversarial, we call Moritz intransigent. In this case, when Moritz quits, he thinks the verdict is wrong, but simply throws in the towel. But with less polarized debate—when Lones is not so biased—he softens his stance, and Moritz’ incentives fundamentally switch. We call him ambivalent if he actually prefers to lose out to the strongest types of Lones who concedes next period. This form of debate is more constructive and focused on learning—for if Moritz quits, then he agrees with the verdict, and the conversation secures a meeting of the minds. The debate genre may change over time, and may differ across jurors. The strategic structure of debate depends on the genre. Strategies in the standard war of attrition are strategic substitutes—when one player concedes more slowly, his rival gains less from holding out, and so concedes faster. This well describes equilibrium incentives here with intransigent jurors. But with ambivalent jurors, ours is instead a game of strategic complements, in which doggedness begets doggedness: For as Moritz grows more partisan, conceding more slowly, Lones learns less from each delay, and fewer of his types concede. Consistent with this, Proposition 1 finds that debate is always ambivalent when jurors are not too biased. Conversely, Proposition 3 shows that communicative debate by sufficiently patient jurors quickly settles into intransigence. Intransigence also obtains in the continuous time limit of our game. For sharper predictions about the impact of juror bias or waiting costs, Propositions 4–7 restrict to low juror biases or assume communicative debate has gone on for a while. In these cases, the debate lengthens as bias grows or waiting costs fall. To see their impact on decision errors, assume that jurors are not too biased, so that debate is ambivalent. In this case, a more hawkish Lones pushes harder for convictions, and thereby reduces errors of impunity. Less obviously, his tougher stance elicits so much pushback from the dove Moritz that the chance of a miscarriage of justice also falls. Here we see the impact of strategic complements with ambivalence: For Moritz can afford to worry less about errors of impunity when Lones grows more assertive, and so can push more for acquittal. But reflecting our strategic dichotomy, at some point, the tables turn: If jurors grow too biased or patient, debate becomes intransigent. Actions then become strategic substitutes, and so Moritz softens when Lones grows tougher. In this case, a more hawkish Lones still limits errors of impunity, but also leads to more miscarriages of justice; however, if jurors are symmetric and both grow more biased, then both decision errors fall. This theoretical insight has applied implications—$$e.g.$$ it intimates as to why one might wish to limit the number of peremptory juror challenges, for a more balanced and polarized jury best determines the truth. An advantage of our model is that it is identifiable: Since our predictions vary in the bias and cost parameters, one can identify them from observables. Our as-if model yields a key intuitive feature of debate: Jurors may eventually play devil’s advocate: Lones might well acquit if he could decide the verdict unilaterally as a dictator, and yet persist in voting to convict. As Lones the debater pays a deliberation cost, one might think him more eager to concede than the dictator, who can end the debate; however, seconding a proposal ends the game, whereas holding out retains the option value of conceding later in light of new information about Moritz’s type. Option value is an important element of dynamic debate, and offers a key contrast with static committee models. We prove in Proposition 8 that devil’s advocacy always arises, as long as jurors are not too biased and delay is not too costly. Our article is technically innovative in many ways. Equilibria are characterized by a possibly infinite sequence of thresholds obeying a non-linear second-order difference equation. Since we know of no general method of solving such a dynamical system, we develop new methods to establish existence and uniqueness. Our existence theorem exploits the Jordan curve theorem as a generalization of the intermediate value theorem. Our uniqueness theorem and comparative statics critically exploit an assumption that jurors’ types have a log-concave density. Many signal distributions obey this condition, adapted from Smith et al. (2016). This condition disciplines the best response functions, since each updates from a truncated signal. We also recursively apply monotone methods to show that the dynamical system is saddle point stable, and our equilibrium loosely resembles a balanced growth path, familiar to macroeconomists. Related Literatures. The cheap talk literature started by Crawford and Sobel (1982) also explores communication by informed individuals before an action. Our model specifically assumes that jurors must agree on a verdict—a motion “moved” and then “seconded” seals the verdict. Another key difference is that our communication is not free. The possibility of trading off the chance to achieve one’s favourite verdict for the extra delay costs overturns a key insight of the cheap talk literature: namely, greater bias leads to worse decisions. There, greater bias renders communication less transparent, and thereby inflates decision errors. In stark contrast, we find that slightly partisan jurors arrive at the truth more often than unbiased jurors. As seen in the dynamic cheap talk literature, Forges (1990), Aumann and Hart (2003), and Krishna and Morgan (2004), dynamic communication can convey more information than static communication. As seen in Goltsman et al. (2009), dynamic communication formally allows an informed agent to credibly commit to send a mixed signal. In contrast, we explicitly model the optimal level of communication when any communications are costly; further, the willingness to persevere is a credible signal of the strength of one’s signal. The committee decision literature, surveyed in Li and Suen (2009), explores free information transmission before a vote.1 Adding some structure, Li et al. (2001) (LRS) allow jurors to cast multiple votes. Our model with vanishing delay costs approximates LRS, but our focal communicative equilibrium has no counter-part in their model. The limit with vanishing delay costs differs from zero delay costs, for costs can be amplified endogenously in equilibrium. Another strand of the committee literature focuses instead on public information acquisition. In this research thread, Chan et al. (forthcoming) [CLSY] is the closest work to us: They consider a war of attrition by voters with ordinally aligned preferences who observe a continuous time public information process.2 In contrast, our jurors have already witnessed the trial evidence, our panel has researched a shuttle explosion, or our committee has seen and read an assistant professor’s research. For instance, FDA panels do not convene until Phase 1–3 trials have ended, and a new drug application is received. We analyse the subsequent deliberation when debaters are thus endowed with their private signals, and seek to learn about each other’s signals. CLSY explore the majority requirements for the vote—a moot point for our jury of two. With unanimity, their war of attrition analysis is akin to our intransigent debate: behaviour by the pivotal voters exhibits strategic substitutes, since a tougher stance by the most hawkish juror induces the most dovish juror to soften. In contrast, our new ambivalent debating genre exhibits strategic complements, and has no counterpart in their analysis.3$$^{,}$$4 Our article contributes to the war of attrition literature. For instance, Gul and Pesendorfer (2012) consider a complete information war of attrition played by two parties with ordinally opposed preferences, and so each seeking to win. In contrast, we arrive at a meeting of the minds with ambivalent debate—when a juror concedes, he is genuinely convinced of his peer’s perspective. In a companion paper Meyer-ter-Vehn et al. (2017), with unbiased and equally patient jurors, any cloture rule that truncates debate at a fixed date lowers welfare. In contrast, asymmetric equilibria in which one juror concedes immediately to his insistent peer, or deadlines that enforce such early agreements, generally increase efficiency when there is a conflict of interest, as in Gul and Lundholm (1995) or Damiano et al. (2012).5 We next introduce and analyse the model, highlighting its novel Bayesian aspects. A two period example gives a foretaste of our results. We prove most results in the Appendix. 2. The Model 2.1. The extensive form game Two jurors $$i=L,M$$, Lones and Moritz, alternately propose in periods $$t=0,1,2,\ldots$$ to convict or acquit, $$\mathcal{C}$$ or $$\mathcal{A}$$, a defendant of a crime. Lones proposes a verdict in period zero. Moritz replies in period one with his own proposal. The game ends if he agrees; otherwise, Lones responds in period two with a proposed verdict, and so on. The game ends when two consecutive verdicts concur—namely, a unanimity rule. A priori, the defendant is equilikely to be guilty or innocent—states $$\theta =\mathcal{G,I}$$. Jurors are partially informed: each has privately observed a signal about the defendant’s guilt. The conditionally iid signals $$\lambda ,\mu\in(0,1)$$ are private beliefs that the defendant is guilty, $$i.e.$$$$\theta=\mathcal{G}$$. Lones is a hawk, hurt weakly more by errors of impunity, and Moritz a dove, hurt more by miscarriages of justice. Jurors’ decision costs are $$1+\beta_L, 1-\beta_M$$ for an actual error of impunity, $$i.e.$$ acquitting the guilty, and $$1-\beta_L, 1+\beta_M$$ for an actual miscarriage of justice, $$i.e.$$ convicting the innocent, where $$\beta_i\ge0$$ is the bias of juror $$i$$. Jurors share the same cardinal preferences over verdicts if $$\beta_L=\beta_M=0$$. We assume $$\beta_L, \beta_M<1$$, ensuring identical ordinal preferences: convict the guilty and acquit the innocent, precluding partisans, who always wish to convict or acquit.6 That $$0<\lambda ,\mu<1$$ and $$\beta_L, \beta_M<1$$ reflects standard jury instructions to “be open-minded”, for in this case, jurors are willing to change their mind given enough evidence. Circuit court jurors are advised: “While you’re discussing the case, don’t hesitate to reexamine your own opinion and change your mind if you become convinced that you were wrong” (Ed Carnes, 2016). The jurors find debate time costly: Juror $$i$$ incurs a waiting cost$$\kappa_i>0$$ per delay period. To avoid trivialities, we assume $$\kappa_i<1-\beta_i$$.7 We say that jurors are symmetric if $$\kappa_L=\kappa_M$$ and $$\beta_L=\beta_M$$. Jurors minimize losses, namely, the expected sum of waiting costs and decision costs. So they are risk-neutral and do not discount future payoffs.8 All told, the game resembles a war of attrition: a stopping game in which each juror trades off the exogenous cost of continuing against the strategic incentives to insist on his preferred verdict; however, this preferred verdict may change as he learns his peer’s type. 2.2. Transforming signals We represent signals as log-likelihood ratios, with different reference states.9 Jurors’ transformed types are $$\ell = \log (\lambda / (1-\lambda)), m = \log ((1-\mu) / \mu)$$. No signal is perfectly revealing, and the common unconditional type density$$f$$ is positive and symmetric $$f(x) \equiv f(-x)$$ of $$x=\ell,m$$, with cdf $$F$$. Since the random signals $$\lambda, \mu$$ are conditionally independent, so too are the transformed (random) types $$\ell,m$$. Let $$p(\ell,m)$$ be the conditional probability of guilt given types $$\ell,m$$. Bayes’ rule implies \begin{equation}\label{life-of-pi} p (\ell ,m) =\frac{e^{\ell-m}}{e^{\ell -m}+1}. \end{equation} (1) Unlike an actual error of impunity, an error of impunity is the ex post event of acquittal despite $$p (\ell ,m)\!>\! \frac{1}{2}$$.10 A miscarriage of justice likewise means that conviction occurs despite $$p (\ell ,m)\!<\! \frac{1}{2}$$. Any type $$y$$ juror entertains the conditional probability density $$f(x|y)$$ that his colleague’s type is $$x$$. Let $$h(x,y)$$ be the unconditional joint density and $$ r(x,y) \equiv h(x,y)/(f(x)f(y))$$ the correlation factor. This yields the conditional density that one’s peer is type $$x$$, given the realized type $$y$$, updating from the common type density: $$f(x|y) = f(x)r(x,y)$$. In Section A.1, we show that $$r(x,y)=2(e^x+e^y)/((1+e^x)(1+e^y))$$, and that $$r(x,y)$$ and $$h(x,y)$$ are log-submodular. For intuitively, since signals about the state are affiliated,11 so too are their log-likelihood ratios; but then the inversely defined random types $$\ell,m$$ are negatively affiliated, as Figure 1a highlights. Figure 1 View largeDownload slide Joint signal density and equilibrium outcomes. (a) Plots contour lines of the negatively affiliated joint type density $$h(\ell,m)$$. (b) Depicts the outcomes of a sincere agreeable monotone strategy profile. Here, $$\mathcal{C}1$$ means conviction in period 1, etc. Lones’ types and (even) cutoffs are on the horizontal axis; Moritz’ types and (odd) cutoffs on the vertical axis Figure 1 View largeDownload slide Joint signal density and equilibrium outcomes. (a) Plots contour lines of the negatively affiliated joint type density $$h(\ell,m)$$. (b) Depicts the outcomes of a sincere agreeable monotone strategy profile. Here, $$\mathcal{C}1$$ means conviction in period 1, etc. Lones’ types and (even) cutoffs are on the horizontal axis; Moritz’ types and (odd) cutoffs on the vertical axis 2.3. Strategies and payoffs Lones’ initial vote fixes the debate roles for the rest of the game. If he initially proposes $$\mathcal C$$, then jurors enter the natural subgame, in which each argues for his natural verdict—$$\mathcal C$$ for Lones and $$\mathcal A$$ for Moritz—until conceding;12 otherwise, they enter the Nixon-China subgame where each argues for his unnatural verdict until conceding.13 Since either subgame is a stopping game, we describe pure strategies by the planned stopping times—the first period a juror plans to concede if the game has not yet ended.14 So Moritz has a strategy described by two (odd) periods in which he first concedes to Lones after either initial proposal, while Lones’ strategy consists of his initial proposal and his planned (even) concession period. We say that Moritz convinces Lones in period $$t$$, say, if Lones quits in period $$t$$. Strategy profiles are equivalent if they imply the same outcome; for any juror’s strategy, call two strategies of the other juror equivalent if the strategy profiles are equivalent. We find Bayes Nash equilibria (BNE), and later prove that any BNE is equivalent to a sequential equilibrium. 3. Preliminary Equilibrium Analysis 3.1. Monotonicity Lones’ initial vote fixes an ordering on types. In the natural subgame, a stronger type of Lones and Moritz is higher, and so more convinced that the proposed verdict is right. In the Nixon-China subgame, stronger types of Lones and Moritz are lower. In a monotone strategy, whenever some juror type holds out until period $$t$$, a stronger type surely holds out until then. Lemma 1. (Single Crossing Property).If a type prefers to hold out from period $$t$$ to $$t'\!>\!t$$, then any stronger type prefers to do so, and strictly so if period $$t$$ is hit with positive probability. Every best response strategy of a juror to any strategy is thus equivalent to a monotone strategy. Stronger types hold out longer as they are not only more convinced of their position, but also more sure that their peer entertains a weaker opposing signal, given their negative correlation. Lemma 1 yields a skimming property of equilibria, familiar in the bargaining literature: Stronger types quit in every period until the end. Moritz’ monotone strategy in the natural subgame is described by a weakly increasing sequence of odd-indexed cutoff types $$(x_t)_{t \in 2\mathbb N + 1}$$, where $$x_{t}$$ is his supremum type that concedes by period $$t$$. Lones’ monotone strategy in the natural subgame is likewise described by a weakly increasing sequence of even-indexed cutoff types $$(x_t)_{t \in 2\mathbb N + 2}$$; monotone strategies in the Nixon-China subgame are described by weakly decreasing sequences $$(x_{-t})_{t \in 2\mathbb N + 1}$$ and $$(x_{-t})_{t \in 2\mathbb N + 2}$$. In other words, type intervals of Lones and Moritz stop in alternating periods; moreover, negative period indexes simply flag that the threshold corresponds to Nixon-China debate. Hereafter, we assume monotone strategies. Consider next period zero. Lones’ strategy is sincere if he proposes to convict when his type indicates guilt, $$i.e.$$$$\ell > x_0$$ for some $$x_0$$, and acquit otherwise. A sincere strategy is responsive if not all types vote for the same verdict, $$i.e.$$$$|x_0|< \infty$$.15 Finally, a strategy of Moritz is agreeable if almost all of his types $$m$$ either plan to second Lones’ initial proposal to convict, namely, $$m<x_1$$, or to second the initial proposal to acquit, $$i.e.$$$$m>x_{-1}$$; this is equivalent to $$x_{-1} \leq x_1$$. Lemma 2. Lones’ best reply to an agreeable, monotone strategy of Moritz is sincere. Conversely, any best reply of Moritz to a sincere, monotone strategy of Lones is equivalent to an agreeable strategy. So up to equivalence, Lones is sincere in equilibrium if and only if Moritz is agreeable. A sincere agreeable responsive equilibrium is characterized by cutoffs $$(x_t)_{t \in \mathbb Z}$$ with $$|x_0|<\infty$$, and:16 \begin{equation} \begin{array}{rcl} -\infty \leq \cdots \leq x_{-3} \leq x_{-1} \!\! &\leq& \!\! x_1 \leq x_{3} \leq \cdots \leq \infty \\ -\infty \leq \cdots \leq x_{-4} \leq x_{-2} \leq \!\! &x_0& \!\! \leq x_2 \leq x_{4} \leq \cdots \leq \infty \end{array}\label{eq-cutoffs} \end{equation} (2) Figure 1b depicts the equilibrium outcomes in this sincerere agreeable case.17 3.2. The propensity to hold out We now characterize equilibrium cutoffs $$(x_t)$$ in terms of indifference conditions. When type $$y$$ of Lones, say, faces type $$x$$ of Moritz, conviction is the correct verdict with chance $$p(y,x)\equiv 1-p(x,y)$$ (recalling (1)) and securing it avoids an actual error of impunity, thereby lowers decision costs by $$1+\beta_L$$; acquittal is the correct verdict with chance $$p(x,y)$$ and securing it avoids an actual miscarriage of justice, and thereby lowers decision costs by $$1-\beta_L$$. Summing up, the net change in expected decision cost from securing Lones’ natural conviction verdict is $$(1+\beta_L)(1-p(x,y))-(1-\beta_L)p(x,y)=1-2p(x,y)+\beta_L$$; an analogous argument applies for Moritz. Using (1), write this expected decision payoff in terms of the type difference$$\delta \equiv x-y$$: \begin{equation}\label{eq-delta} \Delta(\delta, \beta_i)= \frac{1-e^\delta}{1+e^\delta} + \beta_i. \end{equation} (3) As seen in Figure 2a, this payoff falls in the type difference $$\delta$$, since a stronger peer type $$x$$ lowers the chance that the natural verdict is correct. Figure 2 View largeDownload slide The net decision payoff and gap propensity function. Juror $$i$$’s net decision payoff as a function of the gap $$\delta$$ is positive for $$\delta<\underline b_i$$ and negative for $$\delta > \underline b_i$$ — the gain $$DG$$ and loss $$DL$$; the latter includes the delay costs if the peer holds out. Its propensty integral in panel (b) is hump-shaped in the upper gap $$\bar \delta$$, by $$(P3)$$; it vanishes in equilibrium. Panel (a) is numerically simulated for parameters $$\beta_i\!=\!\kappa_i\!=\!0.1$$ and a logistic distribution $$f(x)=e^x/(1+e^x)^2$$ Figure 2 View largeDownload slide The net decision payoff and gap propensity function. Juror $$i$$’s net decision payoff as a function of the gap $$\delta$$ is positive for $$\delta<\underline b_i$$ and negative for $$\delta > \underline b_i$$ — the gain $$DG$$ and loss $$DL$$; the latter includes the delay costs if the peer holds out. Its propensty integral in panel (b) is hump-shaped in the upper gap $$\bar \delta$$, by $$(P3)$$; it vanishes in equilibrium. Panel (a) is numerically simulated for parameters $$\beta_i\!=\!\kappa_i\!=\!0.1$$ and a logistic distribution $$f(x)=e^x/(1+e^x)^2$$ Jurors solve an infinite horizon stopping problem, but it suffices to plan for two periods. Denote the conditional density $$f(x|y,x\ge\underline{x})=f(x|y)/[1-F(\underline{x}|y)]$$. If juror $$i$$ of type $$y$$ believes that his peer’s types $$x\! <\!\underline x$$ have conceded, and that those in $$[\underline x, \bar x]$$ will next do so, then his expected payoff gain in the natural subgame from holding out more period is the propensity function: \begin{equation} \textstyle \Pi_i (\underline x, y, \bar x) \equiv \int_{\underline x}^{\bar x} (\Delta (x-y, \beta_i) - \kappa_i) f(x|y, x \geq \underline x) dx - \int_{\bar x}^{\infty } 2\kappa_i f(x|y, x \geq \underline x) dx. \label{eq-prop-A} \end{equation} (4) The integrand is the net decision payoff, $$i.e.$$ first $$\Delta (x-y, \beta_i) - \kappa_i$$ for $$x\in (\underline x, \bar x)$$ and then jumping to $$-2\kappa_i$$ for $$x\ge \bar{x}$$. It measures the net benefits of an immediate concession in the next period, conditional on the peer type $$x$$. It includes a decision payoff gain for weak conceding peer types $$x$$ with $$\Delta (x-y, \beta_i)>0$$, and a decision payoff loss for strong conceding peer types $$x$$ with $$\Delta (x-y, \beta_i)<0$$—both net of one period delay costs of $$\kappa_i$$. The two period delay cost $$2\kappa_i$$ is incurred if the peer does not concede. Similarly, for the Nixon-China subgame, define the propensity to hold out $$\hat \Pi_i(\underline x,y ,\bar x)$$ by juror $$i$$ of type $$y$$ when types $$x > \bar x$$ have already conceded and types $$x \in [\underline x, \bar x]$$ will next concede: \begin{equation} \textstyle \hat \Pi_i (\underline x,y ,\bar x) \equiv \int_{\underline x}^{\bar x} (- \Delta(x-y, \beta_i) - \kappa_i) f(x|y, x \leq \bar x)dx - \int_{-\infty}^{\underline x} 2\kappa_i f(x|y, x \leq \bar x)dx. \label{eq-prop-C} \end{equation} (5) In contrast to (4), the expected decision payoff now pertains to securing one’s unnatural verdict, and hence has the opposite sign, $$-\Delta(x-y, \beta_i)$$. Indifference by cutoff types $$x_t$$ and $$x_{-t}$$ requires: \begin{equation} \Pi_{i(t)} (x_{t-1}, x_t, x_{t+1}) = 0 = \hat \Pi_{i(t)} (x_{-(t+1)}, x_{-t}, x_{-(t-1)}) \label{eq-indiff} \end{equation} (6) for $$t=1,2,\ldots$$, where $$i(t)\!=\!L$$ for even periods $$t$$, and $$i(t)\!=\!M$$ for odd $$t$$. This is the discrete first-order condition between periods $$t-1$$ and $$t+1$$ for the (omitted) infinite horizon Bellman value. When Lones employs an initial cutoff type $$x_0$$ in a sincere agreeable responsive strategy profile, and plans to concede in period two, should Moritz hold out in period one, using his next cutoffs $$x_{1}$$ and $$x_{-1}$$, then Lones’ initial propensity to convict equals: \begin{equation}\textstyle \bar \Pi_L (x_{-1}, \ell, x_{1}) \equiv \int_{-\infty}^{x_{-1}} \kappa_L f(m|\ell) dm + \int_{x_{-1}}^{x_{1}} \Delta (m-\ell, \beta_L) f(m|\ell) dm - \int_{x_{1}}^{\infty} \kappa_L f(m|\ell) dm. \label{eq-prop-0} \end{equation} (7) Indeed, any type $$m \leq x_{-1}$$ of Moritz immediately agrees in the natural subgame, but holds out in period one of the Nixon-China subgame. Both initial proposals lead to a conviction, but proposing to convict reduces Lones’ waiting costs by $$\kappa_L$$. Types $$m \in (x_{-1}, x_{1})$$ of Moritz agree at once in both subgames, whereupon Lones’ proposal fixes the verdict. In this case, Lones’ decision payoff from proposing to convict equals $$\Delta(m-\ell, \beta_L)$$. Finally, types $$m \geq x_{1}$$ of Moritz agree immediately in the Nixon-China subgame, but hold out in period one of the natural subgame. Hence, both initial proposals lead to the same acquittal verdict, but proposing to convict raises waiting costs by $$\kappa_L$$. Lones’ type $$m = x_0$$ is indifferent between initially voting convict or acquit (and conceding immediately if Moritz does not agree) if: \begin{equation} \bar \Pi_L (x_{-1}, x_0, x_{1}) = 0. \label{eq-indiff-0} \end{equation} (8) We next show that indifference conditions (6) and (8) characterize equilibrium. Cutoffs are tight if whenever all types of one juror concede, all remaining types of the other juror thereafter hold out forever, $$i.e.$$ if $$x_t = \infty$$ at some odd period $$t$$, say, then $$x_{t'} = x_{t-1}$$ for even $$t' > t$$. Theorem 1. (Characterization).A sincere agreeable responsive equilibrium is equivalent to tight cutoffs $$(x_t)$$ that obey monotonicity (2), indifference (6), and (8) if finite, and $$|x_0|<\infty$$. Conversely, any such cutoffs define a sincere agreeable responsive equilibrium. 3.3. Equilibrium existence For insight into equilibria, assume that Lones insists on his initial vote—acquit for $$\ell < x_0$$ and convict for $$\ell > x_0$$—forever after. Since Moritz cannot affect the verdict, he defers at once. This strategy profile is a Bayes-Nash equilibrium for suitable $$x_0$$. It corresponds to an asymmetric outcome of a standard, private value war of attrition, $$e.g.$$Riley (1980). But deference can arise in any period $$t$$ in our game. Had we assumed private juror values (over verdicts), a strategy profile in which, say, Lones surely concedes in period $$t$$ unravels: For (1) no type of Moritz concedes in period $$t - 1$$, and so (2) all remaining types of Lones concede in period $$t-2$$, and so on. But in our common values setting, step (1) of this unraveling logic breaks down: weak types of Moritz do not want win the debate against the remaining strong types of Lones in period $$t-1$$ and hence concede; this in turn gives Lones an incentive to hold out in period $$t-2$$. A $$(\sigma,\tau)$$-equilibrium is a minimal pair of drop-dead dates$$1\le \sigma,\tau\le\infty$$ such that debate ends by period $$\sigma$$ of the Nixon-China subgame and $$\tau$$ of the natural subgame. This equilibrium is deferential in the Nixon-China subgame if $$\sigma<\infty$$, and otherwise communicative. It is deferential in the natural subgame if $$\tau<\infty$$, and otherwise communicative.18 It is deferential if $$\sigma,\tau<\infty$$, and communicative if $$\sigma=\tau=\infty$$. Theorem 2. (Existence).A $$(\sigma, \tau)$$-equilibrium exists for all integers $$(\sigma, \tau)$$, with $$1\le \sigma,\tau\le\infty$$. The drop-dead dates of a deferential equilibrium are enforced strategically in our open-ended game. But one can also view it as the longest possible equilibrium in a truncated game where drop-dead dates are enforced by protocol or regulation, such as cloture rules in a parliament. By Theorem 1, equilibrium cutoff vectors $$(x_t)$$ are described by a second-order difference equation, solving (6) and (8); deferential equilibria also obey the boundary conditions $$x_{-\sigma} = -\infty$$, $$x_{\tau}=\infty$$, and the communicative equilibrium obeys transversality conditions. But there is no general existence or uniqueness methodology for non-linear second-order difference equations. Our existence proof in Section A.5 is intrinsically topological, while our uniqueness proof for small bias, namely, the argument for Theorem 5 in Section A.10, uses monotonicity methods. 3.4. Equilibrium stability We next ask which equilibria in Theorem 2 obey stronger and more robust solution concepts. Theorem 3. (Sequentiality).The communicative equilibrium is a sequential equilibrium. Any deferential equilibrium is a sequential equilibrium. Proof In a communicative equilibrium, all information sets are reached on path, and so Bayes’ rule determines beliefs; the resulting assessment therefore constitutes a sequential equilibrium. Next consider a deferential equilibrium, say with Lones conceding in periods $$\sigma,\tau < \infty$$. If he unexpectedly holds out in period $$\tau$$, then any beliefs over his random type $$\ell$$ derive from some sequence of completely mixed strategies in a sequential equilibrium; that is, the consistency requirement has no bite. For Moritz may interpret the failure to concede as a tremble of weak types; formally, his type $$m$$ may believe that Lones’ type $$\ell$$ obeys $$\Delta(\ell - m) > \kappa_M$$ almost surely. With such beliefs, and expecting that Lones is about to concede, Moritz wishes to hold out. ǁ Yet deferential equilibria do represent a communication failure. Lones concedes not because he is convinced that Moritz is right, but because Moritz refuses to concede. This could not happen if Moritz had to interpret off-path behaviour by Lones as a signal of strength, rather than as a mistake. Inspired by a definition for finite games in Cho (1987), we say that a sequential equilibrium obeys forward induction if either juror who observes a deviation from the equilibrium path must assign probability zero to any types of his peer for whom the observed deviation is not sequentially rational for equilibrium beliefs and any conjecture about future play. Theorem 4. (Stability).The communicative equilibrium satisfies forward induction. If jurors are equally patient, then (1) equilibria that are deferential in the natural subgame violate forward induction if and only if the biases $$\beta_L,\beta_M\ge 0$$ are both sufficiently small; and (2) equilibria that are deferential in the Nixon-China subgame violate forward induction. Forward induction has no bite in a communicative equilibrium, as every information set is hit on the equilibrium path. Next consider a deferential equilibrium, say, with Lones conceding in period $$\tau$$ of the natural subgame. If he unexpectedly holds out, then Moritz must blame this deviation either on Lones’ bias or his information. If Lones is very biased, then Moritz need not infer that Lones holds compelling information for guilt, and Moritz may insist on acquitting. But if Lones is not too biased, then only strong types profit from holding out; forward induction obliges Moritz to acknowledge Lones’ strong information. Still, we can rationalize Moritz’ insistence on acquitting if Moritz is very biased. But otherwise, his weakest remaining type is sufficiently convinced of guilt that he concedes, and the equilibrium unravels. All told, forward induction prunes deferential equilibria when neither juror is too biased.19$$^,$$20 Stability prunes deferential equilibria in the Nixon-China subgame: Since jurors argue against their bias, suddenly holding out for the unnatural verdict betrays strong information. Theorem 4 selects the communicative equilibrium when jurors are not too biased. This choice also follows from reputational concerns by the logic of Abreu and Gul (2000). For assume that not only can either juror be a rational type $$\ell$$, but with a small chance, he is a “behavioural type” who adamantly never changes his verdict. This prunes any deferential equilibria with early drop-dead dates. Indeed, if Moritz expects all rational types of Lones to concede by period $$\tau$$, and Lones instead holds out at period $$\tau$$, then Moritz in period $$\tau + 1$$ infers that Lones is the behavioural type; Moritz then concedes at once, undermining Lones’ deference in period $$\tau$$.21 Besides failing stability, deferential equilibria are actively discouraged in legal settings. Standard juror instructions remind them of their duty to deliberate—for instance: “While you’re discussing the case, $$\ldots$$ don’t give up your honest beliefs just because others think differently or because you simply want to get the case over with” (Ed Carnes, 2016). 3.5. Preliminary analysis for the characterization results 3.5.1. Distributional assumptions Subsequent results also require a type density restriction: $$(\star)$$ The type density $$f$$ is log-concave and has a bounded hazard rate. Log-concavity is satisfied by many standard distributions, and implies a monotone hazard rate $$f/(1-F)$$. A bounded hazard rate is less standard, but holds for the logistic and Laplace distributions.22 Since the hazard rate is monotone and bounded, it finitely converges, say to $$\gamma^{-1}<\infty$$. The inverse $$\gamma$$ of this tail hazard rate measures the thickness of the tail of the jurors’ type distribution, and intuitively is a measure of signal informativeness. 3.5.2. Propensity function properties When peer types $$x\! <\!\underline x$$ of a juror $$y$$ have conceded, and types in $$[\underline x, \bar x]$$ next concede, we call the lower gap$$\underline \delta \equiv y - \underline x$$ and the upper gap$$\bar \delta \equiv \bar x - y$$. Analogous to (4), we define a gap propensity function $$\pi_i (\underline \delta, y, \bar \delta) \equiv \Pi_i (y - \underline \delta, y, y + \bar \delta)$$, obeying: \begin{equation}\label{pi-Pi} \pi_i (\underline \delta, y, \bar \delta) = \int_{-\underline \delta}^{\bar \delta} (\Delta(\delta,\beta_i) - \kappa_i) f(y+\delta|y,\delta \geq -\underline \delta)d\delta - \int_{\bar \delta}^{\infty} 2\kappa_i f(y+\delta|y,\delta \geq -\underline \delta)d\delta. \end{equation} (9) We now partition the gaps into intervals with endpoints $$\underline b_i<b_i<\bar b_i$$, given the bias $$\beta_i$$. These are the respective roots of the net decision payoff and decision payoff, $$\Delta(\underline b_i, \beta_i) - \kappa_i=0=\Delta(b_i, \beta_i)$$, and the crossing point of the integrands in the two integrals of (9), or, $$\Delta(\bar b_i, \beta_i) = - \kappa_i$$: \begin{equation}\label{eq-b} \underline b_i \equiv \log \frac{1+\beta_i-\kappa_i}{1-\beta_i+\kappa_i} \qquad \text{ and } \qquad b_i \equiv \log \frac{1+\beta_i}{1-\beta_i} \qquad \text{ and } \qquad \bar b_i \equiv \log \frac{1+\beta_i+\kappa_i}{1-\beta_i-\kappa_i}. \end{equation} (10) In other words, juror $$i$$ is indifferent between verdicts if his peer’s type exceeds his own by $$b_i$$; he is willing to wait an extra period to achieve his natural verdict if this type difference is $$\underline b_i$$; and he is willing to wait an extra period to get his unnatural verdict if the type difference is $$\bar b_i$$. $$(P1)$$The gap propensity $$\pi_i$$ quasi-increases23in the lower gap $$\underline \delta$$, and is negative for $$\underline \delta < -\underline b_i$$. $$(P2)$$The gap propensity $$\pi_i$$ quasi-increases in the type $$y$$; $$(P3)$$The gap propensity $$\pi_i$$ is hump-shaped in the upper gap $$\bar \delta$$, with maximum at $$\bar \delta = \bar b_i$$. $$(P4)$$The gap propensity $$\pi_i$$ increases in the bias $$\beta_i$$, and decreases in the waiting cost $$\kappa_i$$. To see why property $$(P1)$$ holds, consider Figure 2a. For larger gaps $$\underline \delta$$ with $$-\underline \delta<\underline b_i$$, the gap propensity $$\pi_i$$ grows in $$\underline \delta$$; intuitively, as $$\underline \delta$$ rises, more weak peer types concede, and the willingness to concede rises. For all smaller lower gaps $$\underline \delta$$ with $$-\underline \delta>\underline b_i$$ (not pictured), the gap propensity $$\pi_i<0$$. So $$\pi_i$$ either increases or is negative, $$i.e.$$ it quasi-increases (proof in Section A.4). Property $$(P2)$$ intuitively follows because a higher type shifts probability of the negatively affiliated peer’s type left towards weaker types and the positive area $$DG$$ (proved in Section A.4). Next, the proof of property $$(P3)$$ considers three cases, with the last one pictured in Figure 2a. For upper gaps $$\bar \delta < \underline b_i$$, the positive area decision gain $$DG$$ rises in $$\bar \delta$$. For upper gaps $$\bar \delta \in [\underline b_i, \bar b_i]$$, the negative decision loss area $$DL$$ falls in $$\bar \delta$$, and so the propensity rises. Finally, for upper gaps $$\bar \delta > \bar b_i$$, the negative area $$DL$$ rises in $$\bar \delta$$. This proves $$(P3)$$. Property $$(P4)$$ simply follows from (9). Our article sometimes focuses on the debate when types are large. Here, we have a diagonal monotonicity property with a directional derivative flavour: $$(P5)$$For large enough $$y$$, there is $$\varepsilon > 0$$ with $$\partial \pi_i /\partial\underline \delta \! >\! (1+\varepsilon)|\partial \pi_i/\partial\bar \delta|$$ when $$\pi_i(\underline \delta, y, \bar \delta) = 0$$. Proved in Section A.4, this asserts that the propensity is strictly more sensitive to its first than third argument. For these partial derivatives are proportional to the density $$f(y + \delta|y, \delta \geq - \underline \delta)$$ at $$\delta = - \underline \delta,\bar \delta$$, and this density falls exponentially over the (boundedly positive) interval $$[y-\underline \delta, y + \bar \delta]$$. To analyse the limit game, define the limit propensity$$\pi^{\infty}_i(\underline \delta, \bar \delta) \equiv \lim_{y \rightarrow \infty} \pi_i(\underline \delta, y, \bar \delta)$$. We show in Section A.4 that $$\pi^{\infty}_i$$ exists and inherits the derivative properties $$(P1)$$ and $$(P3)$$–$$(P5)$$ of $$\pi_i$$. 4. Two Period Debate: An Illustrative Equilibrium As a foretaste of our general theory, we explore the deferential equilibrium with drop-dead dates $$\sigma=1$$ and $$\tau=2$$. Here, only a single rebuttal is possible, in which Moritz may push for acquittal against an initial conviction proposal by Lones. So Lones and then Moritz propose verdicts, and there is a conviction if both propose it.24 Since both jurors have the power to acquit unilaterally while conviction requires consensus, this roughly captures a standard presumption of innocence. By Lemmas 1–2, jurors follow sincere, agreeable cutoff rules, depicted in Figure 3a. Lones first proposes his natural verdict (convict) if his type $$\ell$$ exceeds a threshold $$x_0$$, and Moritz opts for his natural verdict to acquit if his type $$m$$ exceeds some threshold $$x_1$$. A lower threshold corresponds to a tougher Lones (or Moritz)—as more types propose their natural verdict. Figure 3 View largeDownload slide Signals: (a) Jury verdicts and (b) reaction curves. Panel (a) depicts jurors’ behaviour, debate outcomes and the decision errors given cutoffs $$x_0, x_1$$. Acquittal is strictly optimal for $$m>\ell$$ and conviction for $$\ell>m$$. We shade the type triangles yielding errors of impunity (EI) and miscarriages of justice (MJ). Panel (b) plots jurors’ reaction curves that fix cutoffs $$x_0, x_1$$. As Moritz grows more dovish or patient, his reaction curve shifts down from $$\Pi_M= 0$$ to $$\Pi_M' = 0$$, as he grows more willing to hold out. Panel (b) is numerically simulated for the signal density and parameters in Figure 2 Figure 3 View largeDownload slide Signals: (a) Jury verdicts and (b) reaction curves. Panel (a) depicts jurors’ behaviour, debate outcomes and the decision errors given cutoffs $$x_0, x_1$$. Acquittal is strictly optimal for $$m>\ell$$ and conviction for $$\ell>m$$. We shade the type triangles yielding errors of impunity (EI) and miscarriages of justice (MJ). Panel (b) plots jurors’ reaction curves that fix cutoffs $$x_0, x_1$$. As Moritz grows more dovish or patient, his reaction curve shifts down from $$\Pi_M= 0$$ to $$\Pi_M' = 0$$, as he grows more willing to hold out. Panel (b) is numerically simulated for the signal density and parameters in Figure 2 4.1. Reaction curves The equilibrium cutoffs $$x_0$$ and $$x_1$$ obey the indifference conditions $$\bar \Pi_L(x_{-1},x_0,x_1)\!=\!0$$ and $$\Pi_M(x_0,x_1,x_2) \!=\! 0$$. Here, $$x_{-1} = -\infty$$, for all Moritz’ types second Lones if he proposes acquittal, and $$x_2 = \infty$$ since all Lones’ types concede to Moritz in period $$t=2$$. Moritz’ reaction curve $$\Pi_M(x_0,x_1,\infty) = 0$$ implicitly yields $$m=x_1$$ monotonically increasing in $$x_0$$. For if Lones hardens his stance by reducing $$x_0$$, his conviction proposals are weaker guilt signals. Moritz responds with a tougher stance, $$i.e.$$ a lower acquittal threshold $$x_1$$. So Moritz’ reaction curve has strategic complements. In contrast, Lones’ (inverse) reaction curve $$\bar \Pi_L(-\infty,x_0,x_1) \!=\! 0$$ is “U-shaped” in Figure 3b. To see this, assume first that Moritz’ cutoff type $$x_1$$ is low. So Moritz acts tough, usually insisting on acquittal, except for very low types $$m < x_1$$. Here, a conviction proposal by Lones usually delays the verdict. In this case, if Moritz further hardens, by reducing $$x_1$$, Lones grows less willing to propose conviction; he softens, raising his cutoff type $$x_0$$. Here, Lones’ reaction curve exhibits strategic substitutes (lower branch of $$\bar \Pi_L \!=\! 0$$ in Figure 3b). At the extreme, when Moritz nearly always asks to acquit (very low $$x_1$$), Lones invariably succumbs to Moritz’ doggedness, even if Lones is convinced of guilt, since a conviction proposal almost never has any impact. Assume next that Moritz’ cutoff type $$x_1$$ is high, corresponding to a soft stance. Then the prospect of a miscarriage of justice is large, and the error of impunity is less likely. Then a tougher stance by Moritz (lower $$x_1$$) reduces Lones’ costs of miscarriages of justice, invoking a more hawkish reply (lower $$x_0$$). So Lones’ reaction curve exhibits strategic complements. Figure 3b plots the reaction curves $$\Pi_M=0$$ and $$\bar \Pi_L=0$$. Moritz’ reaction curve rises with slope less than one,25 by Properties $$(P1)$$ and $$(P2)$$. Meanwhile, Lones’ (inverse) reaction curve, with $$x_0$$ as a function of $$x_1$$, is “U-shaped” with slope less than one in the upward-sloping branch, by Properties $$(\bar P2)$$ and $$(\bar P3)$$ in Section A.4. Hence, the resulting equilibrium is unique. 4.2. Comparative statics As he grows more biased or patient, Moritz’ propensity to hold out for acquittal rises, and so he adopts a tougher stance. Lones’ response depends on whether he exhibits strategic substitutes or complements. If Moritz is initially patient or biased, he acts tough, and usually proposes acquittal. If he grows even more dovish, Lones reacts by softening; for a conviction proposal now simply delays the inevitable acquittal. To wit, Lones’ equilibrium cutoff shifts right. His submissive reaction mode is the hallmark of strategic substitutes. In contrast, if Moritz is initially quite impatient or unbiased, he pushes only weakly for acquittal (a high cutoff $$x_1$$). In this case, if Moritz grows more patient or biased and proposes acquittal more often, Lones can worry less about miscarriages of justice. He then optimally pushes more strongly for conviction. That greater toughness by Moritz begets a tougher reply by Lones is the signature of strategic complements—as depicted in the upper branch in Figure 3b. Next, assume Lones grows more biased or patient. He toughens his stance, proposing to convict more. Seeing a weaker signal in the conviction proposal, Moritz pushes to acquit more. That a tougher Lones begets a tougher reply by Moritz reflects his global strategic complements. 4.3. How delay changes Agreement is delayed when Moritz counters a conviction proposal by Lones with an insistence on acquittal. The chance of this event is the probability mass of types northeast of $$(x_0,x_1)$$ in Figure 3a. If Lones exhibits strategic complements, greater bias or patience of either juror leads each to harden his stance. So cutoffs $$x_0,x_1$$ both fall, and delay unambiguously increases. Next assume that Lones exhibits strategic substitutes. If Moritz grows more biased or patient and thereby hardens his stance, Lones softens. All told, $$x_0$$ rises, and $$x_1$$ falls. With enough bias, Lones’ reaction curve grows infinitely elastic (lower portion of Figure 3b), and $$x_0$$ rises so much that delay falls. In our focal longer, open-ended equilibria, Propositions 4 and 6 find that delay rises as either juror grows more biased or more patient. 4.4. How decision costs change With our presumption of innocence in this truncated equilibrium, errors of impunity are unbounded—the ex post log-likelihood ratio of guilt given an acquittal is unbounded. But consider the maximal log-likelihood ratio of innocence conditional on a convict verdict, $$\delta_1 \equiv x_1-x_0$$; in Figure 3b this corresponds to the (horizontal) distance of the intersection of jurors’ reaction curves from the 45 degree diagonal (not pictured). Assume that Moritz grows either more dovish or patient. The miscarriage of justice measure falls, as he is more willing to fight such mistakes. If Lones exhibits strategic complements, then he toughens in response. This blunts but cannot reverse the effect of Moritz’ tougher stance. But if Lones exhibits strategic substitutes, then he instead softens, further reducing miscarriages of justice. Next consider either a more hawkish or patient Lones. He pushes harder for conviction, increasing miscarriages of justice. Since Moritz’ reaction curve has strategic complements, he toughens in response; this blunts but does not reverse the increase in miscarriages of justice. In our longer, open-ended equilibria, a more balanced story emerges, since debate need not end in period $$\tau =2$$. As Proposition 5 shows later on, if a juror grows more biased or patient, and neither is too biased, then his peer pushes back enough that losses from both errors fall. 4.5. Devil’s advocate Finally, to flesh out the nature of dynamic debate, we contrast Lones’ equilibrium behaviour and his choice were he to call the verdict unilaterally—specifically in the $$\sigma=\tau=1$$ deferential equilibrium. Since proposing conviction risks delay, but not with a unilateral decision, one might think that Lones is less inclined to propose conviction in equilibrium. But for small costs the opposite occurs. Namely, Lones becomes a devil’s advocate, arguing for conviction—despite not wishing that his vote be the last word. For in the deferential equilibrium with $$\sigma=1$$ and $$\tau=2$$, if Lones proposes acquittal, he seals the verdict, whereas a conviction proposal retains the option to concede when Moritz pushes for acquittal. As long as this option value exceeds the delay costs, Lones pushes against his immediate best interests. Playing the devil’s advocate reflects the option value arising in any multi-stage debate setting. 5. Ambivalent and Intransigent Debate 5.1. Two genres of debate Equilibrium balances the costs and benefits of further debate. These include decision gains from winning the debate when winning is ex post optimal, decision costs from winning the debate when losing is ex post optimal, and explicit delay costs. A juror, say Moritz, is intransigent in period $$t$$ of the natural debate if—given Lones’ strongest type who concedes in the next period $$x_{t+1}$$—Moritz’ weakest remaining type $$x_t$$ prefers an immediate acquittal over a conviction one period later: $$\Delta (x_{t+1}-x_t, \beta_i) + \kappa_i > 0$$. This mimics the incentives in a standard, private value war of attrition, in which players wish to prevail against any opponent’s type. We turn to our more novel genre of debate. For debate need not be win-lose. Call Moritz ambivalent in period $$t$$ if—given Lones’ strongest conceding type next period—he prefers a conviction next period over an immediate acquittal, $$i.e.$$ if the expected decision payoff of his natural verdict plus one period’s waiting cost is negative: $$\Delta (x_{t+1}-x_{t}, \beta_i) + \kappa_i < 0$$. When Moritz is ambivalent, he is of two minds, keenly aware that his vote may be a mistake. In this taxonomy, Moritz is naturally intransigent or ambivalent in a period. By (10) and our definitions of intransigence and ambivalence, a juror is \begin{equation}\label{ambiv-intrans} \text{juror}~ i\text{(}t\text{)}~ is \begin{cases} {\textit{intransigent in period t} \text{ if} ~x_{t+1} < x_t + \bar b_i,} \\ {\textit{ambivalent in period t} \text{ if}~ x_{t+1} \ge x_t + \bar b_i.} \end{cases} \end{equation} (11) Put differently, a juror, say Moritz, is intransigent in period $$t$$ when few strong types of his peer Lones concede in period $$t+1$$, for then $$x_{t+1} - x_{t} < \bar b_M$$. But as the marginal conceding type $$x_{t+1}$$ of Lones grows, Moritz transitions into ambivalence, where $$x_{t+1} - x_{t} >\bar b_M$$. We now reformulate our debating genres in terms of the jurors’ best reply functions. If, say, Moritz is intransigent in some period, then a tougher stance by Lones in the next period (so fewer types conceding) reduces Moritz’ propensity to hold out. For it reduces his decision payoff gain from winning the debate against weak types of Lones, by shrinking the positive part of the integral in Figure 4a. This begets a weaker reply by Moritz (greater $$x_t$$)—to wit, local strategic substitutes. Conversely, if Moritz is ambivalent in period $$t$$, a tougher stance by Lones in the next period raises Moritz’ propensity to hold out, by cutting his decision payoff losses from winning the debate against strong types of Lones—the negative integral portion in Figure 4b. This begets a tougher reply by Moritz in period $$t$$; that is, local strategic complements. Figure 4 View largeDownload slide Intransigent and ambivalent debate propensities. With intransigence, in panel (a) above, juror $$i$$’s net decision payoff is positive for conceding peer types $$[x_{t-1},x_{t+1}]$$, before jumping down to $$-2\kappa_i$$. With ambivalence, in panel (b) (and Figure 2a), the net decision payoff is negative for peer types in the interval $$[x_t+\bar{b}_i,x_{t+1}]$$, before a jump to $$-2\kappa_i$$ Figure 4 View largeDownload slide Intransigent and ambivalent debate propensities. With intransigence, in panel (a) above, juror $$i$$’s net decision payoff is positive for conceding peer types $$[x_{t-1},x_{t+1}]$$, before jumping down to $$-2\kappa_i$$. With ambivalence, in panel (b) (and Figure 2a), the net decision payoff is negative for peer types in the interval $$[x_t+\bar{b}_i,x_{t+1}]$$, before a jump to $$-2\kappa_i$$ Figure 5 depicts the two debate genres.26 The disagreement zone is all pairs $$(\ell,m)$$ with $$\ell \leq m + \bar b_M$$ and $$m \leq \ell + \bar b_L$$, $$i.e.$$ where jurors disagree about the best verdict, up to one period’s waiting costs. For intransigent debate in Figure 5a, consecutive cutoff pairs lie inside the disagreement zone. The debate speed, as measured by the staircase step size, is bounded above by the width of the disagreement zone. In contrast, for ambivalent debate in Figure 5b, cutoff pairs zick-zack around the disagreement zone; this bounds the speed of debate from below. Figure 5 View largeDownload slide Intransigent and ambivalent debate: cutoff vectors. Jurors disagree in the disagreement zone between the two indifference lines. These lines collapse onto the diagonal as biases and waiting costs vanish (by (10)). Since only weak types of the rival juror concede in intransigent debate, jurors disagree after the debate; graphically, cutoffs are bracketed by these lines in this case. And because strong types of the rival juror concede in ambivalent debate, jurors agree at the end of debate; in this case, cutoffs straddles two indifference lines Figure 5 View largeDownload slide Intransigent and ambivalent debate: cutoff vectors. Jurors disagree in the disagreement zone between the two indifference lines. These lines collapse onto the diagonal as biases and waiting costs vanish (by (10)). Since only weak types of the rival juror concede in intransigent debate, jurors disagree after the debate; graphically, cutoffs are bracketed by these lines in this case. And because strong types of the rival juror concede in ambivalent debate, jurors agree at the end of debate; in this case, cutoffs straddles two indifference lines While intransigent debate outcomes are win–lose, ambivalent debate can yield inefficient verdicts—$$i.e.$$, convictions in the triangles above the disagreement zone where even Lones prefers acquittals, and acquittals in the triangles below, where even Moritz prefers conviction. We measure these decision errors from an unbiased observer’s perspective. Call the cutoff gap$$\delta_{2t+1}\equiv x_{2t+1} - x_{2t}$$—the log-likelihood ratio of innocence if Moritz’ cutoff type $$x_{2t+1}$$ concedes to Lones’ prior cutoff type $$x_{2t}$$—the (maximal) miscarriage of justice measure in period $$2t+1$$, and the cutoff gap $$\delta_{2t}\equiv x_{2t} - x_{2t-1}$$ is the (maximal) error of impunity measure in period $$2t$$. 5.2. Incidence of debating genres We now explore how the debate genre depends on jurors’ bias, delay costs, and information. The debate genre may differ across jurors and vary over the course of the debate. But we will see that for extreme bias or patience, the genre is unchanging throughout the debate. We say that natural debate is ambivalent (resp. intransigent) if both jurors are ambivalent (resp. intransigent) in all periods of the natural debate; we similarly describe Nixon-China debate. For starters, let Moritz be a partisan who prefers to acquit all defendants, even those whom he knows are guilty: $$\beta_M \geq 1$$ (ruled out). Then, much like in a standard, private value war of attrition, he never wishes that Lones overturn his proposal and natural debate is intransigent. We argue that debate is ambivalent at the opposite extreme, with perfectly aligned interests: no bias $$\beta_L=\beta_M=0$$ and identical delay costs $$\kappa_L=\kappa_M$$. Indeed, for a contradiction, suppose that Moritz’ cutoff type $$x_{t}$$ intransigently wishes to prevail and acquit. With zero delay costs, Lones’ cutoff type $$x_{t+1}$$ conditions on even stronger evidence for innocence and strictly prefers to concede in period $$t+1$$; for Lones only know that Moritz’ type exceeds $$x_{t}$$, since Moritz still remains in the debate. This contradicts the indifference of Lones’ type $$x_{t+1}$$—whence Moritz must have been ambivalent in period $$t$$. This logic persists for all common delay costs $$\kappa_L=\kappa_M>0$$. For in this case, $$\bar b_L = \bar b_M = -\underline b_L = -\underline b_M$$ from (10). If Moritz were intransigent, then (11) implies $$\delta_{t+1} = x_{t+1} - x_{t} < \bar b_M= -\underline b_L$$. By property $$(P1)$$, Lones’ gap propensity $$\pi_L(\delta_{t+1}, x_{t+1}, \delta_{t+2})<0$$ for any $$\delta_{t+2}$$, and his cutoff type $$x_{t+1}$$ strictly wants to concede; contradiction. Lemma 3. (Negative Propensity Proviso).Natural debate is ambivalent provided: \begin{equation}\label{eq-ambturf} \pi_M(\bar b_L, y,\bar \delta)<0 \textit{ and } \pi_L(\bar b_M, y,\bar \delta) < 0 \textit{ for all real } y,\bar \delta. \end{equation} (12) To understand Lemma 3, assume for a contradiction that Moritz, say, is intransigent in period $$t$$; $$i.e.$$ given Lones’ strongest type $$\ell = x_{t+1}$$ who next concede