Equilibrium in Insurance Markets with Asymmetric Information and Adverse Selection Jonathan A. K. Cave Ca MRS ees est eadThe research reported herein was performed pursuant to Grant No. 061B80 from the U.S. Department of Health and Human Services. Washington, D.C. 8-420 SBN 0-8330-0554-5 ‘The Rand Publications Series: The Report is the principal publication doc- umenting and transmitting Rand’s major research findings and final research results. The Rand Note reports other outputs of sponsored research for general distribution. Publications of The Rand Corporation do not neces- sarily reflect the opinions or policies of the sponsors of Rand research, Published by The Rand CorporationR-3015-HHS Equilibrium in Insurance Markets with Asymmetric Information and Adverse Selection Jonathan A. K. Cave April 1984 Prepared under a grant from the U.S. Department of Health and Human SenicesPREFACE This report examines possible outcomes of greater competition in insurance markets. Tt applies directly to the situation where one insurer offers multiple options (e.g., high and low) to an employee group. Many observers argue that requiring employers to make multiple- option insurance available will improve efficiency of the market allocation. Although this report does not directly compare the no-options and multiple-options situations, that is the subject of one of the first analyses of the miltiple-options case. ‘The report removes a rather restrictive assumption in the existing theory of insurance markets and is thereby able to explain observed phenomena in a manner hitherto not possible. Specifically, it describes the nature of insurance offerings in equilibrium if firms offer multiple policies; but it replaces the conventional assumption that each policy must earn nonnegative profits with the more realistic requirement that the portfolio of policies offered by the firm earn nonnegative profits im the aggregate. The equilibrium in general requires that certain policies (low option) subsidize others (high option). Theorems regarding the existence, optimality, and uniqueness of the subsidy equilibrium are presented, together with a simple characterization of the subsidy equilibrium and a comparison with existing equilibrium notions. Because the subsidy patterns, from low to high, that emerge under this formulation appear to characterize multiple-option insurance plans such as the Federal Employees Health Benefits Plan, this model may be more useful than conventional methods in the analysis of such plans. ‘This work was conducted under a grant from the Department of Health and Human Services as part of the Health Insurance Experiment. It is intended for policymakers interested in the welfare effects of competition under conditions of asymmetric information, and for economists and gane theorists concerned with adverse selection problems and games of asymmetric information. The reader should have @ thorough grounding in economics and some familiarity with Nash and perfect equilibria of noncooperative games.SUMMARY In analyzing the effect of adverse selection on insurance markets and other markets where one side of the market may have better information than another, it has been conventional to assume tha contracts are “actuarially fair." In particular, when insurance policies are offered to different types of consumers, and the firm offering the insurance cannot observe the type or risk class of the consumer, it is generally assumed that each contract breaks even on average; there is no subsidization of one class by another. However such subsidies appear to be an important feature of actual insurance offerings. At the sane time, the equilibria that firms reach under the nonsubsidy assumption have some disturbing features. First, if firms assume that their competitors will respond to the presence of a new policy by continuing to offer their current policies, even if they become unprofitable, there may well be no equilibrium. On the other hand, if firms adopt the more sophisticated point of view that new offerings will be greeted with discontinuance of unprofitable offerings, "pooling" equilibria may exist that lump risk classes together. However, under neither set of conjectures about the response of existing firms will the policy offerings necessarily be Pareto-optimal. They will also not deter entry by a sophisticated firm willing to sacrifice actuarial fairness for nonnegative overall profits. We model a process of competition between firms who do not require that all of their policies break even. Instead they will offer portfolios of policies that earn nonnegative profits overall, when portfolios made unprofitable by entry have been withdrawn. It is demonstrated that such behavior always results in an equilibrium for any distribution of the consumer population between a finite number of risk classes. Moreover, except for exceptional cases this equilibrium is unique.To compare these results with the no-subsidy results, it is easiest to think of a world with two lasses of consumers: high-risk and low- risk. Conventionally, if the proportion of high-risk agents is high enough, there will be a separating equilibrium that offers the high- risk agents full fair coverage, and offers the low-risk agents the best fair coverage thet will not attract high-risk agents. This is distinctly inferior to full fair insurance for the low-risk agents, so that there is @ sense in which the low-risk agents suffer a negative externality. If the proportion of low-risk agents is sufficiently high, there will either be no equilibrium (with naive anticipations) or a pooling equilibrium (with sophisticated anticipation). Unlike the separating equilibrium the pooling equilibrium is never Pareto-optimal, but is sensitive to the proportions of high- and low-risk agents Dropping the no-subsidy requirement results in an equilibrium that is always Pareto-superior to the no-subsidy equilibrium, is Pareto-optimal, and is sensitive to the proportions of hight and low risk agents. The low-risk agents subsidize the high-risk agents, and in exchange obtain better terms than they could otherwise achieve. In fact, the subsidy equilibrium is precisely the pair of policies that maximize the utility of the low-risk agents subject to the constraints of zero overall profits; full insurance of high-risk agents; and expected utility maximization by all agents (which separates types) ‘The only exception to this is when the proportion of high-risk agents is so high that this would result in adverse subsidies. For proportions above this critical level (which strictly exceeds the critical proportion dividing separating from pooling equilibria), the subsidy and separating equilibria coincide. The fact that this equilibrium seems to capture a feature of the real world that is absent from the no-subsidy model, and to enjoy the simplicity of a model that predicts the existence of a unique equilibrium as opposed to the multiplicity of "pooling" equilibria that can occur when there are more than two risk classes, suggests that it may be a more realistic model for analyzing the behavior of firms offering multiple-option insurance contractsSection 2 1. IIL. Ww. vy vr. REFERENCES «24.04 + vit - CONTENTS. INTRODUCTION « OPTIMALITY IN A TWO-TYPE MODEL EQUILIBRIUM IN A TWO-TYPE WORLD . PERFECT EQUILIBRIUM .......0044 EQUILIBRIUM AND OPTIMALITY IN A MANY-TYPE MODEL ..+eeeees DISCUSSION AND OPEN QUESTIONS 20 32 44 67 n10 uu. 12. 13. Te 15. 16. v7. 18. 1. 20 at FIGURES ‘The E1-Optimal Policies ‘The "Separating Contracts’ Sone Zero-Profit Pairs ‘The Ends of the Zero-Profit Locus ‘The Zero-Profit Locus An Interior E2 and 3 Optimum Separate B2 and £3 Optima A Contract that Beats (a(P.n),b(P.n)) A Contract that Beats (a*(n),b¥(n)) ..+...+- Equilibriun Selection Anong Multiple E2 Optina Denands with Two Policies Profits with Two Policies Profits with Two Types A Set of Undominated Policies Feasibility in Policy Space Feasibility in Utility Space Optimality in Utility Space Tne Effect of Population Changes E3 Feasibility E3 Corner Optimality Relative Strict Convexity 10 a 12 13 a 5 16 25 28 30 36 35 37 40 49 49 50 51 31I. INTRODUCTION ‘This report examines some of the consequences of allowing firms that operate in markets characterized by asymmetric information between sellers and buyers to compete, using coordinated portfolios of offers that may involve subsidization. Equilibria of markets with adverse selection have been investigated by many authors. The pioneering paper of Rothschild and Stiglitz (RS) described a model in which consumers know their own risk classes (accident probabilities), and firms do not. Their firms are constrained to offer single-option policies, and free entry forces each firm to make nero expected profits; hence, each policy breaks even or is “actuarially fair" for the group it serves. The model is solved by Nash equilibrium between the producers and expected utility maximization by the consumers. For the case of two types of consumers (high- and low-risk), the authors described a distinguished pair of contracts, independent of the population parameters, that would form the Nash equilibrium if one existed. However, they also showed that if the proportion of high risk agents was sufficiently low, no Nash equilibria would exist ‘The RS construction, including the definition of the "separating contracts" that form the Nash equilibrium when it exists, can easily be generalized to arbitrary finite numbers of risk classes, provided the different risk classes can be completely ordered (in terms of expected cost at any policy) and that their utility functions have the "single~ crossing" property: For any two risk classes, and any two indifference curves, one for each class, there is at most one point where the curves cross. However, as the number of types increases, the existence of a Nash equilibrium becomes in some sense less likely. Indeed, Riley showed that in the limiting case of a continuum of types, there is never a Nash equilibrium of the sort described by RS. Within the context of the one-policy-per-firm model, there have been tuo "resolutions" of the RS nonexistence problem. First, Wilson described a different solution concept, termed by him "anticipatory equilibrium," in which the usual Nash assumption that the strategies of‘the other firms will not be affected by one’s oun action is replaced by the assumption that policies which become unprofitable in the face of entry Will be dropped. In the single-policy-per-firm context, this is equivalent to the assunption that firms expect that rival firms on whom they inflict losses will go out of business, An equilibrium relative to these conjectures is a situation in which each fir makes zero profits (actuarial fairness due to free entry), and in which any profitable entry Would become unprofitable once the reactions anticipated by the entrant have occurred. With this definition, equilibrium always exists. It coincides with the RS equilibrium when the latter exists, and otherwise represents a "pooling" equilibrium, in which several different types purchase the same policy. It is important to realize that the Wilson equilibrium can also be described as a Nash equilibrium, albeit of a slightly different gane. Suppose that we construct an extensive-form game corresponding to the story just given: Starting from some initial point, firms will offer policies, observe the results, and enter and exit, until steady state is reached, at which point (in this timeless world) they collect their payoffs. In the RS story, it may well be the case that no branch of this tree (which we shall assume has perfect information) ever terminates; this corresponds to the absence of equilibrium.* If there was an RS equilibrium, then some branches would terminate, and they would do so in an “endgame” of the following stylized form: incumbents pick . entrants zero-profit policies decide Tt should be remarked that we can treat the possibility of infinite branches by assigning to them zero payoffs for all players. In the ison model, the situation is the same, except that there is one more move of the incumbents, and it is of @ special nature: They decide only whether or not to remain in the market. They cannot adjust ¥We could equally well describe a simultaneous-move "endgame’ between incumbent and entrant that did not have @ pure-strategy equilibrium.v -3- their policies, for if they did so we would essentially have returned to the RS world with the roles of incumbent and entrant reversed, and there would be no equilibrium. The stylized endgame is now: incumbents ... entrants ... incumbents pick pick exit or stay A way to describe this game in normal form is to specify a two part strategy for the incumbents. A strategy now consists of two things: @ policy to offer, and a reaction to any pair consisting of incumbent policies and entry policies. This reaction may assume two values: the previous policy (stay) or no policy (exit). Backwards induction will allow us to find the (perfect) equilibria, and it is evident that the reaction part of the incumbent strategy mst correspond to the Wilson anticipations. It follows that the Wilson equilibrium is merely perfect equilibrium in a richer strategy space. Another approach to the RS nonexistence problem is contained in a recent paper by Dasgupta and Maskin (DM), and represents the usual solution to the failure of Nash equilibria in pure strategies to exist: mixed strategy Nash equilibria, DM show that such equilibria always exist for the "adverse selection game," and also that these mixed- strategy equilibria possess some cross-subsidization properties. For our purposes, we merely note that they also represent Nash equilibria of a game with a wider strategy space than that allowed in the RS model. ALL of these models retain the single-option requirement implicit in the constraint that each equilibrium policy offers zero expected profits. RS considered the possibility of offering multiple options, but found that it did not obviate the nonexistence problem. Indeed, as we shall see, it exacerbates the problem somewhat. Wilson did not consider multiple options, and it is natural to ask what the market would look like if firms offered multiple-option insurance, but retained anticipatory conjectures a la Wilson. This is the equilibrium concept that is the subject of this report. It too involves broadening the RS strategy space, but in a different way. Firms are now allowed to offercoordinated portfolios of policies that may involve explicit cross~ subsidization. This idea was first developed in another context by Miyazaki, and applied to insurance markets by Spence, although his definition differs slightly from ours in the case of many risk classes This would seem to be a natural extension for two reasons. First, it is undeniably the case that cross-subsidization does exist (e.g., between Blue Cross High and Low Options) between distinct policies, and is not merely inplicit between different risk classes purchasing the same policy as in the pooling equilibrium. Second, the fact that insurance offered to different risk classes introduces an externality between the profits of various policies suggests an analogy with the theory of joint products, according to which subsidization is entirely reasonable Spence's analysis was mainly devoted to shoving that such policies result in @ solution to 4 particular optimization problem, first identified in this context by RS as a benchmark of efficiency, and to exanining other measures of efficiency. In order to obtain the existence of an equilibrium supporting this optimum, Spence was forced to resort to a modified version of the conjecture entertained by firms in Wilson's model: that firms will react to entry by dropping unprofitable policies. The simple statement that firms drop unprofitable policies is obviously inconsistent with subsidization, so some modification was necessary. In essence, players should anticipate that firms which experience losses as a result of entry will go out of business. This is precisely the conjecture we shall use. It is much simpler than an assumption based on dropping unprofitable policies could be if it were to allow subsidization, and it has certain conceptual advantages as well. If firms are to be allowed to drop only the unprofitable portions of their portfolios after entry, two questions naturally arise: * Why did they not have this freedom before entry occurred? * If they are allowed to alter (rather than delete) portions of their portfolios after entry, why should they not be allowed to alter the terms of the policies?Tf we allow firms either of these freedoms, we return to the RS world, with one important difference: The domain of existence of RS equilibrium is even smaller than it was before, since the RS policies, if they do not solve the optimal-subsidy problem that we shall describe below, can always be attacked by a Pareto-superior, strictly profitable pair of mutually subsidizing policies. As far as efficiency is concerned, the optimization problem that is solved by the Spence-Miyazaki (Si) equilibrium in the two-type case is that of maximizing the utility of the low-risk types subject to the constraints of free choice by each type, zero overall profits (free entry), and the additional constraint that the high-risk types receive at least their full fair insurance utility. ‘The latter constraint rules out perverse subsidies from high- to low-risk types. This problem was, first defined by RS in their discussion of the suboptimality of the RS separating equilibrium. We can compare this with the problem of maximizing the utility of the low-risk types subject to the first two constraints alone, which would be appropriate to a regulated insurance market with controlled entry. If we refer to the latter as full optimality and the former as constrained optimality, we may summarize the results for the twortype case as follows: When RS equilibrium exists in the single-policy world, it may or may not be constrained optimal; the Wilson equilibrium in the single-policy world is only constrained optimal if it is the RS equilibrium as well; and there is a uunigue population distribution for which either is fully optimal. With nultiple-policy strategies, using Nash equilibrium in Gane I, the RS equilibrium, when it exists, is always constrained optimal; the critical population distribution at which it ceases to exist is precisely the distribution for which it is fully optimal; and the policies that characterize it are the seme as those characterizing constrained optimal RS equilibria. If we play according to Game IT, we obtain an equilibrium that coincides with the Gane I equilibrium when the latter exists; exists for all distributions; and is always fully optimal for those distributions having no Gane T equilibrium. In the intermediate range between the distribution at which the RS equilibrium is fully optimal and that at which it ceases to exist in the single-policy world, the Gane IT equilibrium Pareto dominates the RS equilibrium.OUTLINE OF THE STUDY Section II below describes the optimal subsidy problem for the two- type case, and characterizes full and constrained optima, Section IIT addresses the strategic model in the two-type, multiple-option world, providing existence, uniqueness, and optimality results. Section IV discusses the perfectness properties of the equilibria defined in Sec. IIT; and Sec. V extends the analysis to the case of an arbitrary finite number of types. Finally, Sec. VI contains a discussion of the limitations of this model and suggestions for further extension.Il, OPTIMALITY IN A TWO-TYPE MODEL, We are investigating the following situation: There are n, agents of type 1 and n, agents of type 2. For each agent there are two possible states of nature, so the commodity space we deal with consists of pairs of state-contingent incones, (a,b), where a is interpreted a: the agent's income if an accident occurs and b his incone if no acciden occurs. All agents have the sane initial endowment, which we denote (agsby), and the same strictly increasing, strictly concave, indirec utility function, U, for income, An agent of type i has a probability t, of having on accident. For concreteness, we assune t, > ty, so an agent of type 1 will be called a high-risk agent, while an agent of type 2 is a low-risk agent. State-contingent incone allocations (a,b) are evaluated according to the agents’ von Neunann-Norgenstern expected utility function: @ U,(a,b) = ty Ula) + (1 = £00) As a matter of convenience, we make the following definitions: n= nj/(, +n)) nt, + (= nt, tay + (1 = t)by Considerably greater generality is possible, in terms of differing endowments and differing utility functions, but it will not materially affect the results unless it ceases to be the case that representative utility curves of different types intersect in at most one place, and that at @ point of intersection, the indifference curve corresponding to the higher-risk agent is flatter than that of the lower-risk agent.We shall define three concepts of efficiency to use as benchmarks in this model. The first, which we denote El, represents efficiency in a full-information world, where the only constraints are individual rationality and the requirement that the overall portfolio break even. Maximize WU, (a,,b,) + (1 - w)U,(@,,by) subject to: (Q) U,(a;,b,) 2 Uy Cay4b9), all i, and (44) n(tyay + GL ~ ty)b,) + G~ A) (tyay + GL - t,yb,) S x E2 represents the sort of efficiency one might hope for in @ world of asymmetric information; it differs from El by the presence of an additional constraint which specifies that no type prefers, over its own policy, the policy designed for the other type. In practice, this constraint will only be binding on the high-risk types. EZ: Maximize wW,(a,,b,) + (1 - w)U,(a,,b,) subject to: G) U,(@,,0,) 2 Uylaysby), all 4, G44) m(eyay + GQ Db, + = A)CEya, + 1 ~ t,)b,) Sx, and (Ai) 04 (a,5b,) 2 U,Cay5b,) Finally, E3 is the efficiency concept appropriate to a competitive world of asymmetric information. In such a world it is easy to show (as we do formally in Sec. III) that the high-risk agents must achieve at least their full fair-insurance utility U(tyay + (1 - t,)by). If not, then policies exist that would be strictly preferred by the high-risk agents and would earn positive profits regardless of which group or groups purchased them. This is the optimization problem defined by RS (11.3) and used in their discussion of efficiency,E3: Maximize w,(a,,b,) + (1 w)U,(ay,b,) subject to: Gi) U; (a; bj) 2 0; (@y,b9), all i, (4) n(eyay + GL - t,)b,) + C= ad(eya, + (1 - t,)b,) Sx, (iif) U,(ay.b,) 1) 2 U,(ay by), and (iv) U (4,46) 2 Uleyag + 1 ~ £4 )b9) Definition: ‘The portfolio [(a,,,),(a),b,)] is E,-optimal, for i = 1,2,3, if it solves problen E, It is evident that these efficiency concepts are increasingly restrictive. It is also true that efficiency implies that the high- risk agents receive full insurance. Proposition 2.1: Let [(a,,b,),(a,,b))] be E,-optimal Then a, = b, and, if 4 = 1, PROOF: ‘The Lagrangeans corresponding to each problem have the following symmetry property: Whenever a term involving a, appears, it is in the form +, f(a,) for some function f, There is always @ corresponding term of the form (1 - t,)f(b,) and vice versa. This means that differentiating the chosen Lagrangean w.r.t. a, gives the sane first-order condition as differentiation w.r.t. by; and since all terms involving a,(b,) are either Linear in a, or are multiples of U(a,)(U(b,)), we conclude that U"(a,) = U"(d,) 1 and thus by concavity of U that a) = by. The same argument shows that a, = b, in problem £1. QED In Light of this proposition, we shall maintain the following 2 change of notation for the balance of this section:io. Examination also shows that unless t, = t,, £2 end £3 imply a, # by, So that B2 and £3 optimality give smaller velues of the objective function than £1 optinality. It may be useful to give a geometric construction showing the solutions to the optimization problems. There are many representations of insurance contracts; we find it most useful to work directly in terns of the state-contingent net allocations. Thus, the commodity space we shall work with is @ subset of R*. In fact, with initial allocation (agsby), it is unnecessary to consider any allocation with a coordinate greater than ag + by: Figure 1 represents the initial situation, and the solution to problem El ‘The diagram shows three reakeven lines": The set of policies that would break even if purchased by the high-risk types is denoted bel, and defined by tya + (1 - t,)b = tyay + (1 - t,)bp} the set of policies that would break even if purchased by the low-risk types is denoted be2, and defined by tya + (1 - t,)b = tyay + (1 = ty)bg; the set of policies that would break even if purchased by (a random sample of) ‘the "market" population is denoted bell and defined by ta + (1 - t)b= tay + (1 = t)by, The El-optimal policy for type i is located at the a be2 452 = (00,ba) Accident payott Novaceident payott Fig. 1 -- The El-optimal policies-u- intersection of bei with the full-insurance line defined by a = bs this is where U, is maximized on bei. ‘The E1 optima are achieved at unique policies. Since they involve full insurance for both types, the E1 problem can be written: maximize w(e) + (1 - w)U(a) st ne + (1 = -n)a =x, where a is the full-insurance outcome for the low-risk players. The uniqueness of @ and ¢ follows by concavity: When we move to the asymmetric-information setting, the separation constraint (iii) comes into play. First, we can define the E2-optimal contracts for the two types that lie on their respective breakeven lines. As RS showed, this pair of contracts, which comprise their separating contracts, consist of the El-optimal high-risk policy, denoted by C, together with the policy on be2 that maximizes U, subject to separation. This latter is denoted R in Fig. 2. a5 (a6, be Accident payott Us No-accident payoft Fig. 2 The "separating contracts"Be This pair of contracts can be used as an initial feasible solution to problem £2, We shall now construct the locus of all feasible solutions to E2 that might be optimal. This entails the constraints (i)-(iii) above, together with a, = b,, in light of Proposition 2.1. The locus we want is that of contracts for the low-risk types which would make zero overall profits when offered in conjunction with a riskless contract that is equally ranked by the high-risk types. Clearly, R is on this locus, when offered in conjunction with G. Another point on this locus is the intersection of the market-odds line be with the full- insurance locus. Now pick any point Bon bell. If this were purchased by both types, it would make zero profits. Tf the high-risk types were sold any contract on a line through B parallel to bel, and the low-risk types were sold any contract on a line through B parallel to be2, that combination would also make zero profits. (See Fig. 3.) Conversely, given any zero-profit pair of contracts [(21,0,),(aj,b))], if we draw a Line parallel to bel through (a,,b,)s and a line parallel to be2 through (a,,b)), the Lines will intersect betf at the same point. 45° = (a0, bd) Accident payoff No-accident payott Fig. 3 Some zero-profit pairsa5 In light of the individual rationality constraint (4) above, we shall start the construction using that riskless policy that gives the high-risk types their initial utility U,(a),by). This is denoted by Cy in Fig. 4, In conjunction with this policy, the low-risk types must be offered a policy along the line labeled Ly. By quasi-concavity of U, there are at most two points on Ly that satisfy the separation constraint (iii). Of these, the "lover" one, with a b. This, combined with the steeper slope of the low-risk indifference curves, means that they would prefer the policy offered to the high-risk types. Since this policy lies above beNl, its purchase by both types is incompatible with zero profits. Accident payoff No accident payott Fig. 4 => The ends of the zero-profit locusue We can therefore generate the entire zero-profit locus in the following manner: Beginning with any point E on the full-insurance line between Cy and A, we first find the point F where the line through E parallel to bel intersects bel. Let us denote by L(E) the line through F parallel to be2; the policy for the low-risk types, G, is located at the (lower) intersection of L(E) and the indifference curve of the high- risk types passing through E. The entire locus, labelled IL,, is shown in Fig. 5. Several things should be clear from this construction. First, if we limit our attention to points below A as argued above, then the separation constraint only affects the high-risk types. Second, the locus lies above bet and intersects be2 only at R. Third, the slope of the locus can be determined directly from the constraints Using Proposition 2.1, we get. Gi) ne + (1 = ny(tga + (1 - ty yb) = x, (444) U(e) = tC) + G - EDUC) Accident payoff @= (as, be) No-accident payott Fig. 5 -- The zero-profit locus<5 Total differentiation gives Gi") nde + = n)(tgda + (1 = ty)€b) = 0 (4ii") U'(e)de = t,U' (adda + (1 - t,)U" ab ‘The slope of the locus is therefore: [n(2-t,)U"(b) + Ca-n) (1-t,30"(e)] (2) da/eb = = ———_—_____. [n(e,)0"(a) + G-n)(e,)0"(e)} Since the market will direct our attention to the special case w= 0, we shall concentrate on it for the balance of the section. There are several possibilities for the solution of the optimization problems £2 and E3. Under certain circumstances, the solution will be an interior maximum of the utility of the low-risk types among the contracts on the zero-profit set. If such an interior maximum occurs between A and R, then it is @ solution to both the E2 and £3 problems (see Fig. 6) 1£ such an interior maximum falls between R and D, then it is E2-optimal but not E3-optimal. The reason is that it violates constraint (iv), More interesting is the observation that such an optinum involves “perverse subsidies" in the sense that the insurance firm would earn profits on its high-risk clientele which would pay for the losses it suffered on the low-risk policies. This is the reason why Accident payott No-accident payoff Fig. 6 -- An interior E2 and E3 optimumeo such an optimum would not be expected to prevail in a market, In such a case the E3 optimum would be R. This is shown in Fig. 7 A corner solution at A is impossible, since the zero-profit locus and the market-odds Line be are tangent at A. However, since this conmon slope is flatter than that of be2, it is flatter than a low- risk indifference curve, which has slope -(1 - t,)/t, along the full- insurance Line It ds possible to continue the zero-profit locus beyond D, and in principle one might expect the possibility of optima located beyond D. However, this would violate the individual-rationality constraint (i) for the high-risk types. In case there is an interior maximum, a useful necessary condition can be obtained by equating the slope of a low-risk indifference curve to that of the zero-profit locus found above. This gives: (C-njt,(1-t,) 1 1 @) ——— = vent U"(b) 45° Accident payott No-accident payott Separate E2 and E3 optimaoe As the population parameter n varies, the positions of the optima will vary as well. We shall demonstrate that the solution to E2 is unique and varies continuously with n, and that there is a well-defined “critical value" n* such that the solution to E2 solves E3 if n < n* On the other hand, if n > m*, the solution to E3 will be R, while the solution to £2 will involve perverse subsidies, and will lie between R and D. Proposition 2.2: The solution to E2 is unique in the two-types case, if U is twice continuously differentiable and strictly concave Tt ds also a continuous function of the population parameter n. PROOF: We can write E2 as: max ty Ula) + (1 = ty) U(b) s.t. ULC = G+ ny tga + C= ty )b))/n] 2 ty Ola) + G = ty) D0) The I0C's defining the solutions to this give the optimality condition, Eq. (3). If multiple solutions to Eq. (3) exist, then they must give the same utility to the type 2 agents (we are only interested in the uniqueness of the global solution to the problem, not whether other local solutions might exist). Thus, there must exist (a,b) and (a',b') satisfying Eq. (3) and GH) ty Ula) + C2 = ty) UCb) = ty UCa) + G1 - £,) UC") In particular, if we use this condition to write b implicitly as a function of a, then we want to show that the LHS of the above equation is monotone in a. Let us define the following functions: Vox) = 1/0") “) WO = U"GD/(" G))?ao We can now write the LHS of Eq. (3) as U'(c)[V(a) - V(b)], and its derivative (wort. a) as (6) U'(e)[V' (a) = V'Cbyb] + U"Ce)e"[V(a) - V(b)] However, a V(x) = WOO so the quantity in (6) can be written: @ U'(e)[WObyb" = Wa] # UNCeDe" (Va) = VOB] We can evaluate b’ and c' by taking the total derivatives of Eq. (4) and of the constraint in E2: @ bi = -[ty/ - t,)11V0)/V(@)] (a9) ct [(e, - €,)/G ~ t,)11VCCI/VC@)] 1 Substitution of Eqs. (9) and (10) in Expression (8) and division by U'(c) > 0 gives (ee-t,)1 1) W(e)[———] 1 Vey/¥Ea)]- WO) [a-t,)] a-t,) 1EV@b)/VCad] = WCa) But strict concavity implies that W(x) is uniformly negative, and moreover since a V(a), so Expression (11) is positive in each term. Hence the RHS of Eq. (3) is strictly monotone in a and there exists but a single global solution to E2. Moreover, the LHS of Eq. (3) is monotone-decreasing and continuous in n, so this unique solution is also continuous in a. OED In solving £3, if constraint (iv) is binding, we are at R, and we may specify the contracts completely using the following versions of the constraints:-19- (12) G(R) tyay + (1 - tybg 13) ty (aR) - = (1 = £5) (by = BD) 44) UCe(R)) = £,U(AR)) + A ~ EUR) None of these conditions depends on n, so (a(R),b(R),¢(R)) is independent of n, as is the RHS of Eq. (3) evaluated at the R-contracts. ‘The LHS is monotone-decreasing and continuous in n, so that n* defined by n¥)t,(I-ty) a 1 as) ———— = ver) ——— - ——__ nA(t ty) v'CaR))—U'(ERD) is unique and has the properties claimed above. 1f U is not strictly concave, there may be multiple optima, and ‘they may vary discontinuously with n. In Sees. TTI and IV, we shall assume that the E2-optimal policy is unique. Proposition 5.2 below provides « generalization of Proposition 2.2 above, with a simpler proof,oe IL, EQUILIBRIUM IN A TWO-TYPE WORLD As noted in the Introduction, the nature of equilibrium in market: with asymmetric information is heavily dependent on the conjectures entertained by firms about the reactions of their rivals and on the strategies employed by firms. All of the models under discussion involve a free-entry condition, which means that whatever the equilibrium set of policies may be, it will be characterized by expected profits of zero for each firm. Under the one-policy-per-firm models of RS and W, this in turn implies that each policy must break even for the group it is intended to serve. In multiple-option models, this means that the portfolio of policies offered by a firm, while it may involv some cross-subsidization, breaks even overall Im essence, these models are models of equilibrium with externalities; since the consumers are not assumed to behave strategically, the asymmetric information merely serves as an externality Linking various policies on the market, and multiple-opt ion policies provide a means of internalizing the externality. ‘The strategic assumptions may be phrased in various ways, but they can be reduced (for the two-type case) to two: no reaction by rivals in terms of the policies offered (which we call Nash conjectures); and no reaction by rivals in terms of strategies, where a s| tesy may make policy offerings contingent on nonnegative profits (which we cal anticipatory conjectures after W). Either type of conjecture can be used to define a conjectural equilibrium in the sense that no firm conjectures that it can increase its profits by a unilateral change of strategy. It is also clear that for every equilibrium in strategies that say, "offer policies p," there is an equivalent equilibrium in strategies that say “offer policies p as long as they earn nonnegative expected profits; otherwise offer no policies." In the present model, we suppose thet individual firms are free to offer as many policies as they wish. A strategy for @ firm is thus a set {(a,,b,)} of policies (state-contingent allocations), together with a statement as to which if any of these policies will be offered shoulda1 ‘they prove to be unprofitable. ‘The nature of this "second" part of the strategy threatens to become unmanageable; in general, we are asking for something like @ strategy which describes the set of policies that will be offered as a function of the spectrum of policies offered by the other firms. While this may be a fruitful way to think of the strategy space, it poses analytical difficulties. Let the space of policies be denoted P; it is @ convex compact subset of R*, If Px is the power set of P, then a strategy can be viewed as 4 function: s(i):P-——9P+ To each n-tuple of strategies, 5, we would like to assign an outcome R(s) belonging to PexP~x ... Px. Denoting this as R(s) of all R(j,5) except R(i,s), we require that: RCs), -.., R(ays), and writing R(-i,s) as the union RGys) = s(4)(R(-i,8)) for each i To each outcome R (J, +++) Rn) we can assign payoffs, so that in principle it is easy to write the normal form of this gane. We shall denote the payoff function thus obtained as H(s). In what follows, we shall confine our attention to some particular special cases of this setup. Game |: The Rothschild-Stiglitz game In this version, each s(i) is @ single element of P; moreover, it does not depend on the policies offered by the other firms. Let HCL, (ai,bi)), (a, b,,) be the profits of firm i when it offers (a,,b,) and its rivals offer (a_,4b_,)- 22 Game II: The Wilson game In this version, each s(i) is again @ single clement of P. I dopends on the policies chosen by other firms in the following manner: Each s(i) has the form: SCU)(RO-4,5)) = (ajsb,) iff HG, (ay4b,),RC-4,8)) 2 0 5 (agsbg) otherwise. Game 111: The multiple-option Nash game In this version, each s(i) is a subset of P; however, it doos not depend on the policies chosen by the other firms. Game IV: The multiple-option anticipatory game In this version, each s(i) is @ subset of P; the dependence on the policies chosen by the other firms is as follows: 8G)(RO-i,8)) = ((agsb,)) iff HG, £(a, sb, RC-4,5)) 2 05 (agsbp)} otherwise. In other words, each firm chooses a fixed set of policies that it will offer if they make nonnegative profits, and offers the null policy otherwise These games differ only in the strategy spaces of the players; the definition of equilibrium is the same for each game: Definition 3.1: An equilibrium is an n-tuple SHC), vey s(n) of strategies with the property that for each i, and each s(i), HCi,s*) 2 WCi,s(4),8*(-4)) where s(-4) = s(1), ---) S(i*1), SUH), ..5 s(n). Although all these games have compact convex spaces of pure strategies, the payoff function is not continuous, so there is no guarantee thet an equilibrium in pure strategies exists. In fact, we shall find that a pure-strategy equilibrium always exists for Games II and IV, but does not always exist for Games I and IIT-23- There are sone simple relations between the various games. An equilibrium in either of the games with Nash conjectures (I or III) gives rise to an equilibrium with an identical outcome in the corresponding game with anticipatory conjectures (II er IV, respectively). Moreover, since the strategy spaces for the single- option games (I and II) are subsets of the multiple-option strategy spaces for the games with corresponding conjectures (III and IV, respectively), an equilibrium in the latter games that happens to involve single-option policies will be an equilibrium for the former games. However, equilibria of the single-option games are not necessarily equilibria of the corresponding multiple-option games. We shall find that there is a close connection between this question and the optimality properties of the single-option equilibria ‘Throughout the balance of this section, we shall focus on the influence of the population parameter, n, on the existence and optimality properties of equilibria. We begin by defining the conditions under which equilibria of Games I and III will exist, and by showing that equilibria for Games II and IV always exist. Proposition 3.2: There exist n~ m~ stems from the fact that a "pooling" policy located on (or below) bell may offer higher utility to the type 2 agents than (a(R),b(R)) does. In this case, such a policy will attract the type 1 agents as well. The existence condition is therefore that bel lie below the type 2 indifference curve through (a(R),b(R)). This means that: ©) Vin) = max tU(a) + (1 = t,)U[t(ag - a/(1 - t) + by) $ t,UCa(R)) + (1 = £,)U(D(R)) Since t = n,, + (1 - n)t, is monotone-increasing in n, V(n) is strictly monotone-decressing in n (és the mérket Line bel approaches the type 1 fair-odds Line bel, the best type 2 utility along that line decreases) Therefore, there is a single, well-defined n~ at which Bq. (5) holds with equality. For n < nv, Bq. (5) does not hold, and the policy (a(P,n),b(P,n)) that solves the problem on the LHS of Eq. (5) will be strictly preferred to the RS policy by both types, and will break even.oo However, this policy cannot be an equilibrium policy (unless n= 0), by virtue of the fact that the low-risk agents have steeper indifference curves than the high-risk agents). Figure 8 shows that there will exist a policy preferred to (@(P,n),b(P,n)) by only the low- risk agents. Since the market fair-odds line is less steep than bel, this policy can be chosen below bel, and will therefore be profitable. Property (ii): The n¥ referred to is that defined in Sec. IT; it is the point at which the separating contracts (c(R),c(R)), (a(R),b(R)) are E2-optimal. First we shall show that n* > n~. Then we shall show that n > n* implies that (c(R),c(R)), (a(R) ,b(R)) are equilibrium contracts. Finally, we shall show that for n~ nx it is useful to recast their definitions. Let us assume that the E2-optimal policies are unique (for example, let U be smooth and strictly concave), and denote then (as functions of n) by (or(n) e*(n)) for the high-risk players, and (a*(n),b*(n)) for the low-risk players 45° erter Accident payott (en) No-accident payott Fig. 8 -- A contract that beats (a(P,n),b(P,n))eee Recall that n* is defined by: (6) [Cor(n*) ,c* *)) (aE), BECMD)] = [ (CCR) ,€(R)), (@CR), BR) and that n~ is defined by: ) Wore) = b,0(a(R)) + CL = B,200DER) By definition, n* > nv if there exist a subsidizing pair o policies that is Pareto-preferred to both (a(P,n~),b(P,n~)) and (e(R),c(R)), (@(R),B(R)) and that earns strictly positive profits at n~. By definition, (8) ,U(alPyne)) + (Lt, )UC(R,nH)) = BAUCA(RD) + A ~ ,)00BER) and since t, > ty, o £,U(a(P,ne)) + C= £,)UCEP,AY)) > UCCCR) Therefore, consider the pair of policies [(c',c'),a(P,n=),b(P,n~) defined by (10) Ue") = EyU(aCPn)) + CL = eIUCCP MD) The losses when the type 1 agents all purchase the policy (c',c') are strictly the minimum losses over all policies giving then ¢ utility of at least t,U(a(P,n~)) + (1 - t,)UCb(P,n~)), Therefore, by continuity U and of Linear functions, there exist positive nunbers © and 6 such that: Gd) nlet +6} + C= mieya(Pyne) + 1 tb nr) +) [Link])) + CL = ty)UC([Link])) , and (13) £,0(a@P,ne) #2) + = £, UCP.) + £) > Ver) This shows that n¥ > nv,<2. Now suppose that n 2m, In this case, we know that (e(R),€(R)), (a(R) ,b(R)) is B3-optimal. Moreover, we know that it is an equilibrium of Game 1, which means that any policy that is profitable when (c(R),c(R)), (a(R) ,b(R)) is on the market must be a separating pair of policies. However, any such pair must necessarily involve perverse subsidies, which in turn means that it must offer the high-risk agents lower utility than U(c(R)). The result of offering such a pair is that only the low-risk agents would purchase it, resulting in losses for the potential entrant. In case n* > n 2 n~ the separating policies (c(R),c(R)),(a(R),b(R)) are not E2 (which coincides with E3) optimal. Since there exists a Pareto-preferred pair of subsidizing policies with zero profits, a continuity argument such as that used above shows that neither the separating policies nor any other pair of policies that are not E2-optimal can be equilibrium policies. However, the E2-optimal policies are also not equilibrium policies, for the same reason that the pooling policies were not equilibrium policies in Game I. At the policy offered to the low-risk agents, their indifference curves are steeper than the indifference curve of the high-risk agents, which runs through both their policy (c*(n),c#(n)) and through the low-risk policy (at(n), there exist policies to the *(n)) by virtue of the separation constraint. It follows that southeast" of (a*(n),b*(n)) that would attract only the low-risk agents. Such poli es, one of which is shown in Fig. 9, must be profitable to offer unless (e*(n),b*(n)) Iies on or above bel, which contradicts n No-accident payott Fig. 9 -> A contract that beats (a*(n),b*(n)) When we turn to anticipatory equilibrium, the nonexistence problem disappears, but the efficiency properties become more interesting Wilson has shown that for Game IT, the pure-strategy equilibrium policies are given by (¢(R),c(R)),(a(R),b(R)) when n m%, we must show that [(c*(n),c#(n)),(a*(n),b*(n))} is a set of policies that could be offered in equilibrium. First, since there does not exist a Pareto-improving pair of policies thet offers nonnegative profits, it follows that any pair of policies capable of earning profits if it serves the entire market will not be purchased by at least one type when it is first offered. In order to induce the incumbent firm to exit, such an attacking pair of policies must attract the incumbent's low-risk customers. In order to do so, the low-risk customers must be offered a policy above the zero-profit locus derived in Sec. II. But then there is no policy for the high-risk types to purchase that they would prefer and that makes nonnegative overall profits in combination with the low-risk policy ep At this point, it is worth remarking that the above results do not depend on the uniqueness of the E2-optimal policy. In case there exist multiple solutions to E2, they must be Pareto-ranked, offering the same utility to the low-risk types and increasing utility to the high-risk types. In this case, the power of the competitive market to select optima is enhanced, in that it will select the Pareto-superior solution to EZ. To see this, suppose that any of the other E2-optimal pairs of policies was being offered. Offering the low-risk agents a policy slightly better than that which they would purchase at the Pareto-superior solution to £2 and offering the high-risk types a policy slightly worse than the policy they would purchase at the Pareto-superior E2 optimum will result in attracting both types and earning positive profits. This is shown in Fig. 10 ‘The equilibrium argument of Proposition 3.4 works equally well when the attacking policy is a pooling policy. We can again sum up the efficiency results for the anticipatory games in a Corollary Corollary 3.5: ‘The pure-strategy equilibrium of Game II is E2-optimal if n= ns E3-optimal if n $m; and neither E2- nor E3-optinal if a > n*,45° Equilibrium optimum Accident payoff Optimum No-accident payoff Fig. 10 -- Equilibrium selection anong multiple E2 optima ‘The pure-strategy equilibrium of Gane IV is: E2-optimel if n 2 nt; and E3-optimal if n < n* The results of this section indicate that the availability of multiple-option insurance will change the nature of market equilibrium, Té firms entertain Nash conjectures, then the set of populations for which the RS equilibrium exists shrinks, since any suboptimal (in the E2 sense) RS separating policy can be attacked by a Pareto-superior, strictly profitable, pair of policies. This will in turn be vulnerable to attack by policies that siphon off the profitable low-risk agents. On the other hand, if firms entertain anticipatory conjectures, the use of mult iple-option policies serves to enhance the welfare properties of the set of policies offered in equilibrium. For the range of populations where the single-option anticipatory equilibrium separates types, the use of multiple options results in an unambiguous reduction in the externality due to asymmetric information, with the gains beinga divided between the two types. For the range of populations where the single-option anticipatory equilibrium pools different types, the use of multiple options will increase the utility of the low-risk types but may decrease the utility of the high-risk types. There is a sense in which the Nash and anticipatory conjectures form opposite ends of a continuum. Suppose that the gane were to be repeated indefinitely, but that strategies could be selected only once. After a firm opened its doors, the only modification of its policies it could make would be to go out of business forever. If firms in such a game (which is an extreme example of what Merschak and Selten (1978) call an "inertia supergane") are completely myopic, in the sense that they completely discount the future, the Nash equilibria of this game will be equilibria of Games I and IIT. On the other hand, if firms evaluate their payoffs according to the limiting (undiscounted) average, then the equilibria of the repeated game will be equilibria of Games 11 or IV (depending on whether firms can offer multiple options). Seen in this way, it seems entirely reasonable that the true state of affairs lies somewhere between the extreme nonexistence of the myopic (Nash) conjectures and the extremely long-range view enbedded in anticipatory conjectures. In any event, it sens that miltiple options must form part of the model, since the existence of subsidization is indisputable.- 32 - IV. PERFECT EQUILIBRIUM In addition to asking whether a particular set of policies would be offered as a result of market equilibrium, it is useful to have some informal fon as to the sensitivity of such predictions to slight mistakes on the part of the players. This raises the issue of perfectness, since a perfect equilibrium is in some sense one that is immune to such smal] mistakes. ‘There are several different ways to define perfect equilibrium, and two of them have sone application here. First, if a strategy is defined over more than one stage, or opportunity to play, then the latter prescriptions of that strategy are contingent on the earlier choices of all players. For example, in Game II or Game IV, the specification that 4 firm faced with losses on its aggregate portfolio would withdraw from the market gives an example of such a contingent prescription or threat In such @ game, one (relatively weak) perfectness requirenent would have it that any such threat must be credible in the sense of being the most profitable available action should the circumstances leading to the threat actually occur This can be seen as requiring robustness to the small mistakes that lead to the invocation of the threat. If an opponent enters by mistake with @ set of policies that inflict losses on the incumbent, it must be nore profitable for the incumbent to exit then to carry on (we persist in the stipulation that no firm is allowed to modify an existing policy). It is fairly easy to see that the threat implicit in the strategies of Games II and IV is credible under this view. Indeed, the threat to continue to offer losing policies implicit in Games I and ITT is incredible. In the two-type case, it is also not a very intelligent threat, since there any entry that is eventually profitable is also immediately profitable. A more subtle perfectness condition emerges when we study the normal form of the game, either by ignoring stages or by collapsing then into single "strategies." An equilibrium is perfect in the normal form Aff it is the limit of a sequence of e-perfect equilibria as ¢So approaches 0, An e-perfect equilibrium is a completely mixed-strategy netuple where no player uses a pure strategy that does not maximize his payoff with probability more then ©. Such strategies are called inferior strategies A strategy is dominated if there is another strategy that pays at least as much no matter what one's opponents choose, and pays strictly more for at least one choice by the opponents. If the opponents are playing completely mixed strategies, a dominated strategy is always inferior. Therefore, dominated strategies are never used in perfect equilibria If there are only two players in the game, this characterization is exact: An equilibrium is perfect iff it gives positive probability only to undominated strategies. If there are more than two players, it is possible for there to exist equilibria in undominated strategies that fail to be perfect. In this game, however, the strategies of the other players affect any one player's payoff only in the aggregate; what matters is only the set of policies offered by competing firms Therefore, the sets of strategies attainable with independent and correlated play by the other players are the same, and the trenbling- hand perfect equilibria are the undominated-strategy equilibria. A vital question in seeing whether the equilibria of the previous section are perfect is that of who the active players are, and what options they have available to then. As we have modeled the game, only the firms are active players. This does not make much difference to the equilibrium analysis, since players are assumed to maximize their expected utility and thus are playing @ best response if they always expect firms to offer the same policies. Indeed, even if firms are expected to go out of business in response to losses, it is never ine consumer's interest to behave otherwise. However, whether the consuners are active players docs make a difference to the behevior of firms, which will be different if they anticipate that consumers will also make mistakes. For example, it is easy to see that any separating equilibrium (which is the only sort that really permits consumer mistakes) can be upset by such mistakes, regardless of whether or not it involvesae subsidization. The reason is that, while it is a large mistake for low- risk agents to purchase the policy intended for high-risk agents, the latter are indifferent between the policies they are supposed to buy and those offered to the low-risk agents. Thus, any firm that offers such a pair of policies will face negative expected profits if the consumers make small mistakes. This might lead to the choice of policies which earn a small positive profit if consumers do not make mistakes, but this is incompatible with equilibrium. This is not really fatal, since it leads to the use of an evequilibrium concept that produces nearly the sane results as before, but it does complicate matters For the balance of this section, we shall concentrate on the case of two firms who are the only active players in a market containing two types of customers. First, let us consider the single-policy model. We take the strategy space to be the set of all nonnegative pairs dominated by ([Link]). If two firms offer policies p and p', we can describe the profits of each as a function of both policies, as in Fig. 11. If the rival firm locates in region A(p), the firm offering p will earn the sane profits as previously. If the rival locates in C(p), the firm offering p will earn no profits. If the rival locates in B(p) Accident payott No-accident payoff Fig. 11 => Demands with two policies- 35 - (D(p)), the Firm offering p will earn the profits (losses) it formerly earned on its low-risk (high-risk) customers alone. In addition to these data, we need to know the position of the breakeven lines to evaluate each of these possibilities To facilitate the discussion, let us denote the profits earned on ‘the high- and low-risk customers by Hi(p) and H2(p), where p = (a,b): HLCP) = alty(ay = a) + C= t,)Cbg = BD] H2(p) = CL = n)[tg(ag ~ a) + GL ~ t,)(by ~ BD We also need to classify the strategy space according to the profits earned on each type, as in Fig. 12. Now we can show that the policy of full fair insurance for the high risk types is undominated. Let this policy be p~ = (c(R),c(R)) (defined in Sec. IT, Eq. (9)), and let p be any alternative policy the firm és considering, Accident payolt No-accident payoff Fig. 12 -- Profits with two policies~ 36 - Tf pis in (p+) then Hy(p) > Hy(p~) and Hy(p) < H,(p~) If p ds in G(p~) then H,(p) < Hy(p*) and Hp(p) < Hy (p~) If p is in J(p>) then H,(p) < Hy(p-) and Hp(p) > Hy (pr) To show that p is undominated, it suffices to demonstrate that for each position p could occupy, there is a place q for the other firm to locate such that the resulting profits to the firm offering p~ strictly exceed those it could earn offering p. For example, if p is in E(p~), then if q belongs to A(p~) and C(p), the profits if the firm offers p~ will be H(p~) + Hp(p~) > 0, while the profits if the firm offers p will be 0, This will always be possible since E(p~) belongs to A(p~). In general, it is useful to classify the possible demands that could result from entry as follows. If p is in: then possible demands for p~ and p (respectively) are: tp) (22,1) 3122) 5 (12,0) § (12512) 5 (0,0) 5 (1,0) 5 (1,1) 5(2,0)5 (2,2) Bip) (12,12) 5 (12,2) 5 (00) 5 (0,295 (1,0) 5 (1,1) 51,2) 5 (1,12) 5 (2,2) c(p) (12,12) 5 (1,12) 5 (1,1) 5 (2, 12) 5 (2,2)5 (0,12) 5 (0,1) 5 (0,2)5(0,0) Dep) (12,12)5 (12,1) 50,0) 50,195 (1,1) 5 (2,0) 5 (2,1) 5 (2,2)5 (2,12) where 1 and 2 stand for types 1 and 2, respectively; 0 stands for the empty set; and 12 stands for both types 1 and 2 There are eight possible combinations of utility (A-D) and relative profitability (E-J). They will be denoted by both letters; the intersection of A(p~) and E(p~) will be denoted AE(p~), etc. If p is in: then demands favoring p that are feasible are: AE(p~) (12,095 (2,0) AF (p>) (12,295 (12,0)5 (2,095 (2,2) Al(p~) (12,0); 12,195 (1,195 (2,0) BG(p~) (22,12)5 012,295 (2,235 G41) BF(p~) (12,295 (2,2) ca(p~) (42,12)5 G13 (2,2) DG(p~) (22,12); 12,2)5 1,195 (2,0); (2,2) DI(p~) 2,1); G45 @,0)Since none of these is empty, it is evident that p~ is undominated. ‘The same cannot be said for (a(R),b(R)), the policy that is offered to the low-risk agents in the RS separating equilibrium. We shall use the technique developed above to determine the entire set of undominated single policies. First, we may divide the policy space into profitability regions relative to the two classes of consumers; Fig. 13 shows this division. ‘There are nine regions, which may be described by the sign of the profits they would yield if policies in them were purchased by the type 1 and type 2 agents, respectively. ‘Thus, (a),by) is in region (0,0); (e(R),e(R)) is in region (0,+); (a(R) ,b(R)) is in region (~,0); and both policies of 4 nonperverse subsidy equilibrium belong to region (-,+). Next, for any candidate policy p~, we must classify the possible positions that a policy which dominates it could occupy. The two important characteristics are the relative profitabilities of the two policies for each group (determined by the categorization E-J), and the relative attractiveness of each policy (determined by the categorization (a0, be) Accident payolt Novaceident payott Fig. 13 -- Profits with two typeseos A-D). The latter is important even if the two policies never compete in the market, since it determines the differential allocation of demand in the face of entry. ‘The possibly nonempty regions and the relative profitabilities of pr and p located in the indicated region relative to p~ are: Region max [I (p) 4H, (P™)] nax(Hy(p) 5H (P~)] aE He H,() ar He Hp) as or) He) A Hye) HW (p) Br He He) 36 Ho He) cE, Hy () Hy) ce 4) yo) or Hy) #@) DE H,@) 4H (P) DG Hye) Hy (>) Dy He) 1 (P) This information can be combined with information as to the location of Pr to determine whether or not there is a q such that p~ can give a strictly higher payoff against q than p can. Suppose, for example, that p~ was in region (-,-), and that p was in AE(p~). We would then have Hy(p) > Hy(p~) < 0 and H,(p) > Hy(p-) < 0 In fact, it is easy to see that p can be chosen such that Hy(p) > 0 > Hy(p-) and Hy (p) > 0 > Hy(p~). Tt is clear that no matter where q is placed, p will afford at least equal profits to p~, and will often provide better profits. Thus p dominates p-. The sane is true of pr in (0,-), (=,0), and (0,0). For the (+,+) region, if p is in E(p~), the profits earned with p~ will be higher if q is in D(p~) and in G(p). This intersection will be nonenpty unless p itself belongs to C(p~), giving both types higher utility, Thus if CE(p~) is empty, no policy in E(p~) dominates p~. This will be the case, if the type 2 indifference curve through p~ is-39- steoper than bel. If p is in G(p~), then any q in E(p~) will belong to E(p), and will result in strictly larger profits for p~. Tf p is in F(p=), entry at q will result in larger profits for p~ if q is in D(p~ and either D(p) or C(p), both of which are nonempty. Finally, if p i in J(p~), then q will give pr higher profits if it is in B(p) and not in A(p~) or if q belongs to both B(p) and C(p~). The former set i nonempty if a(p-) < b(p~); otherwise, the latter set is nonempty. Thus. policies in the (-,+) region are undominated by other single policies a Jong as CE(p+) is nonempty. Similar arguments show that policies in the (4,0), (0,4), and (++ regions are undoninated by single policie Now consider a policy p» in the perverse subsidy region (+,-), and suppose that p belongs to E(p~); then p earns higher profits on both types than pv. The only pattern of demand that necessarily gives p~ higher profits is when p~ is sold to the low-risk types alone, while p is not purchased. Therefore, we would need to find q in both B(p~) an C(p); but these are disjoint sets. lence no policy in this region ij undominated The set of policies undominated by any single policy is thus th set of policies that earn positive profits if purchased by the low risk agents, and which have the property that CE(p~) is noneapty. Thi: latter condition can be interpreted geometrically, since it i equivalent t: ub) tae) — 2 —— ue) t,(1-t) This is the equation of the ray through the origin defined by the a condition that the low-risk indifference curves are tangent to the high: risk breakeven line. This condition is independent of n and of (asby) so that several geometries are possible. A typical set of undominated policies is shown in Fig. 16. Tt remains only to check that none of these policies is dominated by a mixed policy. First consider the policy p~ = (c(R),c(R)). Let p be any mixed policy that dominates p~. It must therefore be the case that p affords nonnegative profits when offered in competition withaie Accident payoff No-accident payoff Fig. 14 -> A set of undominated policies (a(R),b(R)). Since pr is in equilibrium with this policy, everything in the support of p and therefore p itself must give exactly 0 profits, which in turn means that the support of p is contained in the set of policies not preferred to p- by the high-risk types. Now suppose that firm 2 were offering a policy q slightly smaller in both coordinates than p-. By the argument used to show that p~ was undominated by any Pareto-inferior policy (the fact thet it lies in region (0,+)), the support of p mist be confined to the high-risk indifference curve running through p> Cand through (a(R),b(R))). Now consider any policy offered by the entrant that attracts no customers: Clearly, the incumbent's profits are strictly maximized by offering p-. Hence the support of p ust be exactly pr. In goneral, the payoffs if, (p) and ,(p) are strictly decreasing Linear functions of p, and the payoff function H(p,q) can be defined as: 0 if p belongs to A(q) [i.e., q belongs to C(p)] Hy(P) 4f p belongs to D(q) [i.e., 4 belongs to B(p)] (2) Hp.) H,(p) if p belongs to B(q) [i.e., q belongs to D(p)] Hyp) + H)(p) i€ p belongs to C(q) [i-e., q belongs to A(p)]oe Although this payoff function is not convex, it has a convex envelope as P varies anong pure policies, so that if p can be dominated by @ mixed policy, it can be dominated by a pure policy. Hence, the set of undominated single policies is equal to the set of policies that earn positive profits on at least one type. The payoff function is different for Game II, which has anticipatory conjectures. For one thing, each of H,(p), iy(p), and Hy(p) + H,(p) is replaced by the maximun of itself and 0. Since exit is absorbing (no reentry is allowed), we can modify the payoff function according to the signs of Hy(q), ly(q), and H,(q) + 1l,(q)- Denoting the sign-triple pattern as the triple (sgn(H,(q)),sen(Hly(q)), sgn(i(q) + 4,(q))), we have Te Gtstst), (+,0,+), (0,44), or (0,0,0): 0 SE p ds in C(q) max [31,(p),0] 1f p is in D(q) Hepa) = max [3l,(P),0] if p is in BCq) max [H,(p) + H,(p),0] if p is in Aa) IL. (4,4), oF (4,0): 0 if p is in C(q) H(p,q) = max [Hy(p),0] if p is in D(a) max [H,(p) + Hy(p),0] Af p as in A(q) or Bq) HI. Gy4s-) max (H(p),0] if p is in D(q) H@,a) = max [Hj (p) + H,(p),0] otherwise TV. Ges), G20) 0 if p is in Ca) H(pyq) = max [Hp(p),0] Sf p is in B(a) max [Hj (p),0] if p is in A(q) or D(q)~42- Ve Gyr), Osea) max [H)(p),0] if p is in B(q) (a) max [Hj(p) + H,(p),0] otherwise WI. (5247): H(p,a) = max (Hy(@) + Hy(p),0] ‘The analysis of perfectness must take account of this variation in the payoff function, However, for the dominated regions of the game with Nash conjectures, q could be chosen anywhere, so that (c(R),¢(R)) remains undominated and (a(R),a(R)) remains dominated, We shall now see that the Wilson pooling equilibrium is undominated. Since it lies in the "normal subsidy" region, and earns zero profits overall, we may classify it as (-,+,0) according to the above scheme. What we need to show is that it satisfies condition (1). However, at the pooling equilibrium the slope of the low-risk indifference curve equals that of the market breakeven line, which is steeper than the high-risk breakeven line The analysis of perfectness is more complicated for the multiple- option games. To simplify matters, we shall limit our attention to showing that policy pairs arising as interior subsidy equilibria are undominated ‘The reason the analysis is different is that in the miltiple-option model the industry structure is not well specified. In principle, if there are two firms in the market, one would expect both to offer the same pi r of policies. Division of the profits between the two is irrelevant, since any equiproportional allocation gives each firm 2ero profits. However, if we assume that the other firm offers the same pair of policies, we can show that the set under consideration is undominated. Any other pair of policies that would make nonnegative profits if it served the whole market would attract no low-risk customers, since their indifference curve is tangent to the zero-profit locus. Hence, such @ policy would make losses, On the other hand, policy that would make positive profits if offered in competition with