IDENTIFICATION AND ESTIMATION OF A DISCRETE GAME OF COMPLETE INFORMATION P. Bajari, H. Hong, S. P. Ryan Econometrica 2010
28 February 2012
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
1 / 30
Outline of the Talk
1
Introduction
2
The Model
3
Simulation
4
Estimation
5
Identification
6
Conclusion
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
2 / 30
Introduction
Discrete games: games with a finite number of players, moves, events, outcomes, etc. Examples: labor force participation, entry, technology choice, advertising, analyst stock recommendations, etc. Usual representation in the normal form ⇒ For a given set of payoffs, (quite possibly despite refinements) multiple Nash equilibria. ⇒ Problem with Identification
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
3 / 30
Introduction Approaches to deal with multiple Nash equilibria: 1 Equilibrium selection mechanism in which equilibrium is selected as part of an econometric model; e.g. random or extremal eqm selection (Bjorn and Vuong (1984), Jia (2008)) 2 Restricting attention to a particular class of games and identifying of payoff parameters even with multiple equilibria; e.g. entry games (Bresnahan and Reiss (1990,1991), Berry (1992)) 3 Bounds estimation of best-response functions; e.g. pure Nash eqm games (Tamer (2002), Jia (2005), Berry and Reiss (2007) and so on). Most papers in the literature are on entry games. Econometrically, a discrete game is naturally modeled by conditional logit or multinomial probit models.
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
4 / 30
Introduction
This paper studies identification and estimation of discrete games, allows for multiple and mixed strategy equilibria, proposes a simulation-based estimator. Model primitives: players’ utilities an equilibrium selection mechanism for the probability of an equilibrium being played.
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
5 / 30
Introduction
This approach, applicable to any normal form game with complete information (mixed strategies included), explicitly models and estimates equilibrium selection mechanism, proposes conditions for semiparametric identification of structural parameters underlying payoff functions and the equilibrium selection mechanism: Identification at infinity (Index restriction): Rich support for covariates assuming structural utility parameters can be indexed. Identification by exclusion restriction: Covariates that affect one player but not the other locally identify payoffs and equilibrium selection mechanism, e.g. firm-specific cost shifters.
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
6 / 30
The Model Setup
For i = 1, 2, ..., N players, the finite action set is given by Ai A = Xi Ai with a = (a1 , ..., aN ) The vNM utility function, ui : A → R, is given by ui (a, x, θ1 , i ) = fi (x, a; θ1 ) + i (a) where x are covariates, θ1 parameters, a is the action profile and i (a) a random preference shock. A Nash eqm is a vector π = (π1 , ..., πN ) s.t. each agent’s mixed strategy is a best response.
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
7 / 30
The Model Setup
i (a) reflects info about the utility (that is common knowledge to players), but unobserved by the econometrician i (a) are iid Q with Q the density gi (i (a)|θ2 ) and joint distribution g (|θ2 ) = i a∈A gi (i (a)|θ2 ). Let u = (u1 , ..., uN ). Then E (u) denotes the set of Nash equilibria, λ(π; E (u), β) denotes the probability that π ∈ E (u) is selected, and β is the vector of parameters
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
8 / 30
The Model
Including λ allows for estimating how equilibrium is selected, rather than imposing an ad hoc rule For λ to generate a well-defined distribution, ∀u, β, X λ(π; E (u), β) = 1. π∈E (u)
A parsimonious, parametric model of λ, λ(π; E (u), β) = P
exp(β · y (π, u)) 0 π 0 ∈E (u) exp(β · y (π , u))
By estimating β, one can find which equilibrium best matches the outcomes observed in the data. (⇒ Sort of a rationality measure.)
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
9 / 30
The Model
Combining everything together, the probability of observing a certain action profile in one play: ! Z N X Y g ()d. P(a|x, f , λ) = λ(π; E (u(f , ), x), x) π(ai ) i=1
π∈E (u(f ,),x)
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
10 / 30
Simulation Goal: Simulating the model to get the method of simulated moments (MSM) estimator. Simulation of the probability of observing a certain action profile in one play: P(a|x, θ, β) =
Z
X
λ(π; E (u(x, θ1 , )), β)
π∈E (u(x,θ,))
N Y i=1
! πi (ai ) g (|θ2 )d.
⇒ Computationally burdensome! Changing the variable of integration from to u, with the density of u|θ, x being h(u|θ, x) and i ∼ iiN(0, 1), the density becomes YY h(u|θ, x) = φ(ui (a) − fi (x, a; θ1 )|0, σ) i
Y. Emre Ergemen (UC3M)
a∈A
Reading Group
28 February 2012
11 / 30
Simulation
With the change of variable from to u, prob. of a certain action profile in one play: ! Z X N Y P(a|x, θ, β) = λ(π; E (u), β) πi (ai ) h(u|θ, x)du i=1
π∈E (u)
With importance sampling, where q(u|x) is the importance density, P(a|x, θ, β) =
Z X
λ(π; E (u), β)
i=1
π∈E (u)
Y. Emre Ergemen (UC3M)
N Y
Reading Group
! h(u|θ, x) πi (ai ) q(u|x)du q(u|x)
28 February 2012
12 / 30
Simulation
(s)
(s)
For a given value of x, a pseudorandom sequence u (s) = (u1 , ..., uN ), s = 1, ..., S, is drawn from q(u|x). Then P(a|x, θ, β) is simulated as ! N S X h(u (s) |θ, x) Y X 1 ˆ P(a|x, θ, β) = λ(π; E (u (s) ), β) πi (ai ) q(u (s) |x) S (s) s=1
Y. Emre Ergemen (UC3M)
π∈E (u
i=1
)
Reading Group
28 February 2012
13 / 30
Simulation
Advantages of this simulation approach: ˆ P(a|x, θ, β) is an unbiased estimator of P(a|x, θ, β). Using u (s) , θ and β do not enter into E (u (s) ). ⇒ No need to recompute the eqm set with parameter changes Simulator function is smooth in underlying parameters. ⇒ MSM objective function will be well behaved.
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
14 / 30
The Estimator
Maximum simulated likelihood estimator is asymptotically biased unless √ST → ∞.
√
T −consistent but
Maximum simulated moments estimator is consistent and unbiased for any value of S. To form the estimator, enumerate elements of A from k = 1, ..., #A ⇒ #A − 1 linearly independent conditional moments (sum of #A probabilities is 1). ωk (x) : vector of weight functions
At true parameters of the DGP, E [1(at = k) − P(k|x, θ, β)]ωk (x) = 0.
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
15 / 30
The Estimator The sample counterpart of the moment for the k−th element of a vector of #A − 1 moments is T 1 X mk,T (θ, β) = [1(at = k) − P(k|xt , θ, β)]ωk (xt ). T t=1
So, the simulated analog is T 1 X ˆ m ˆ k,T (θ, β) = [1(at = k) − P(k|x t , θ, β)]ωk (xt ). T t=1
Then, for positive-definite weighting matrix WT , the MSM estimator is ˆ β) ˆ = arg min m (θ, ˆ T (θ, β)0 × WT × m ˆ T (θ, β). (θ,β)
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
16 / 30
Identification
Unknown model parameters to be identified: f (a, x), entering into the utility function λ(x), probability of a certain equilibrium being played Two approaches to identification: 1 Sufficient conditions to identify payoffs and the eqm selection mechanism as covariates’ support gets larger 2 Identification based on agent-specific payoff shifters
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
17 / 30
Scale and Location Normalizations
Assumption The payoffs of one action for each agent are fixed at a known constant. Adding a constant to all deterministic payoffs does not perturb the set of equilibria. ⇒ Location normalization.
Assumption The joint distribution, = (i (a)) is independent and known to all agents and the econometrician. Multiplying all deterministic payoffs by a positive constant does not alter the set of equilibria. ⇒ Scale normalization
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
18 / 30
Difficulty of Nonparametric Identification
P(a|x, f , λ) =
Z
X
λ(π; E (u(f , ), x))
N Y i=1
π∈E (u(f ,),x)
! g ()d πi (ai )
Holding x fixed, model primitives to be identified: f (a, x) and λ(x) Denote P(a|x) = H(f (a, x), λ(x)) where H is the map defined above.
Assumption H(.) is continuously differentiable with the Jacobian denoted by DHf ,λ (x)
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
19 / 30
Difficulty of Nonparametric Identification
Global identification: For f 0 (a, x) and λ0 (x) satisfying the P(a|x) equation, if no other pairs of f (a, x) and λ(x) also satisfy it, then (f 0 (a, x), λ0 (x)) is globally identified. Local identification: ∃ Nx (an open neigborhood of (f 0 (a, x), λ0 (x))) s.t. there’s no other ˜ ˜ vector (f˜(a, x), λ(x))) ∈ Nx with (f 0 (a, x), λ0 (x))) 6= (f˜(a, x), λ(x))) that also satisfies P(a|x).
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
20 / 30
Difficulty of Nonparametric Identification
Necessary condition for identification: Full-rank DHf ,λ (x) Sufficient conditions for local identification: ”The rank of DHf ,λ (x) equaling to the number of parameters in (f (a, x), λ(x))) at (f 0 (a, x), λ0 (x))).” (Rothenberg (1971), Thm 6) global identification: ”∃ a square submatrix W of DHf ,λ (x) with dimension (f (a, x), λ(x))) s.t. W has a positive determinant and W + W 0 is positive definite throughout the parameter space.” (Gale and Nikaido, (1965)) ⇒ Checking these conditions in games with multiple players and strategies can be difficult.
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
21 / 30
Identification at Infinity
Covariates have full support Mean utilities are defined by a linear index of the covariates ⇒ fi (a, x) = xia βia for all ai . Two-step identification procedure: Identify mean utilities with covariate values that give a unique eqm with prob close to 1 → Perturb the covariates locally to identify the utility parameters Under invariance assumption of eqm selection mechanism, identify the eqm selection probabilities from observable choice probabilities
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
22 / 30
Identification at Infinity Assumption k
For any i = 1, ..., N and action profile k−i ∈ A−i , there exists a set τ−i−i of covariates x s.t. lim
k
P(a−i = k−i |x) = 1.
kxk→∞,x∈τ−i−i
For each player i, covariates x can be shifted along a dimension s.t. k−i is a dominant strategy.
Assumption ∀ i, a ∈ A s.t. mean utility fi (a, x) is not normalized, ∃ L0 > 0 s.t. a
inf min eigE [xia xia |x ∈ τ−i−i , kxk ≥ L] > 0.
L≥L0
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
23 / 30
Identification at Infinity
Theorem Under those assumptions, βia and hence fi (a, x) is identified up to the ex ante normalization for all i and all a for fixed x. Now, an invariance property to identify the eqm selection probabilites:
Assumption Equilibrium selection probabilities depend only on latent utility indices, ρ(x, ) = ρ(u(a, x, )) and are scale invariant w.r.t the latent utility indices, i.e. ρ is homogeneous of degree 0.
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
24 / 30
Identification at Infinity Finally, to ensure total utility indices to be approximated arbitrarily well:
Assumption There exists a set τ s.t. for all δ > 0, fi (a, x) lim min P − 1 < δ = 1. ui (a, x, ) |x|→∞,x∈τ i,a
Theorem Under all assumptions, equilibrium selection probabilities ρ(u(x, )) are all identified from observed choice probabilities whenever the cardinality of E (u(x, )) is less than or equal to #A − 1.
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
25 / 30
Exclusion Restrictions
Search for variables that influence one equation but not the other, as in sample selection models. Here, covariates that shift the utility of agent i but not uj , for any j 6= i, or the eqm selection mechanism λ.
Assumption For each i, there exists some covariate xi that enters ui and not the others. Furthermore, ρ(u(α, x, )) depends on u(α, x, ) only through a set of sufficient statistics. λ cannot depend freely on utility indices.
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
26 / 30
Exclusion Restrictions
Theorem Under these assumptions, if #xi are sufficiently large, the necessary order condition is satisfied. ⇒ Identification is possible if we have covariates that are indexed by agent’s identity, i. Examples of imposing exclusion restrictions in discrete games: Strategic entry games: Airline operations (a possible shifter: number of connecting routes) (Holmes (2008), Jia (2008)) Technology adoption in the presence of network effects: Adoption of video-conference systems by in-firm employees (possible shifter: level of employees)
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
27 / 30
Application
The application presented in the paper: California Highway Procurement Auctions: Strategic entry by bidders Estimation of the probability of selecting a particular eqm, λ All strategic possibilities allowed Two stages of the games: Bidding and Entry. A possible equilibrium is decided based on the expected entry profits. λ depends on whether joint payoffs are maximized (efficiency).
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
28 / 30
Conclusion
1 2 3 4
Estimation of both utilities and equilibrium selection mechanisms Generally applicable toolbox for normal form games In-depth discussion of identification Theory overlapping with a real-life example
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
29 / 30
Conclusion
Possible directions for future research: Identification and estimation of network games: For example, identification of peer effects on learning, estimating the diffusion of information in general networks, etc. Extensions to continuous games (e.g. Cournot competition) via set-identification tools Discrete games with incomplete information using Bayesian updating
Y. Emre Ergemen (UC3M)
Reading Group
28 February 2012
30 / 30