14.384 Time Series Analysis, Recitation 12 Notes

Review of the Asymptotics of Extremum Estimators 1 14.384 Time Series Analysis Recitation 12 Review of the Asymptotics of Extremum Estimators These notes review GMM, indirect inference, and MLE asymptotics. They discuss consistency, asymptotic normality, and testing. GMM, indirect inference, and MLE all involve minimizing some objective function, so they are collectively referred to as extremum estimators. A good reference for them is the Handbook of Econometrics chapter by Newey and McFadden (1994). I mostly copied this from notes from 385, so there might be references to stuff from lecture that we never covered in this course. Types of Estimators • Moment methods – GMM – IV • Extremum methods – MLE – M-estimators – Quantile regression – Minimum Distance Minimum Distance 2 Figure 1: Relationship among estimators Courtesy Elsevier, Inc., http://www.sciencedirect.com. Used with permission. a a Source: Newey and McFadden (1994) Minimum Distance We covered GMM and MLE in detail in lectures (in 385). Here we will go through minimum distance. θ̂ = arg min fˆn (θ)� Ŵ fˆn (θ) (1) where plim fˆn (θ0 ) = 0. Includes: • GMM with fˆn (θ) = 1 n � g(zi , θ) and Ŵ =  � ∂ ln f (zi |θ) • MLE with fˆn (θ) = n1 and Ŵ = I ∂θ p • Classical Minimum Distance (CMD): fˆn (θ) = π̂ − h(θ) where π̂ → π0 = h(θ0 ). Usually, π are reduced form parameters, θ are structural parameters, and h(θ) is a mapping from the structural parameters to the reduced form. – Example: Chamberlain (1982, HoE 1984) approach to panel data. Model: yit = xit β + ci + eit , E[eit |xi , ci ] = 0 Reduced form: regress yit on all xi· to get πt . h(θ): we know that xi· πt is the best linear predictor of yit given xi· . We also know that BLP (yit |xi· ) =BLP (xit β + eit |xi· ) + BLP (ci |xi· ) =xit β + xi· λ So if we stack the πt into a t × tk matrix π, we know that π = h(β, λ) = IT ⊗ β � + ıT λ� where β is k × 1 and λ is tk × 1. Consistency 3 p • Indirect Inference: is mathematically the same as CMD, fˆn (θ) = π̂ − h(θ) where π̂ → π0 = h(θ0 ), but the justification is slightly different. We have an economic model, which we are not entirely certain is the true DGP (or perhaps is just difficult to compute the likelihood for), but we do believe can capture some important features of the data. These features of the data are summarized by the parameters of an easy to estimate auxillary model, π. h(θ) gives the estimates of the auxillary model that we would expect if our economic model were the true DGP and had parameters θ. h(θ) is often calculated through simulation. – Example: DSGE (taken from 14.384 notes, for a real application see e.g. Del Negro,Schorfheide, Smets and Wouters (2007)) Consider a simple RBC: c−γ − 1 γ s.t. ct + it = Aλt ktα kt+1 = (1 − δ)kt + it maxE0 � ωt λt = ρλt−1 + �t �t ∼ N (0, σ 2 ) This model has many parameters (θ = (ω, γ, A, α, δ, ρ, σ 2 )) and it would be difficult to write down a likelihood or moment functions. Moreover, we don’t really believe that this model is the true DGP and we don’t want to use it to explain all aspects of the data. Instead we just want the model to explain some feature of the data, say the dynamics as captured by VAR coefficients. Also, although it is hard to write the likelihood function for this model, it is fairly easy to simulate the model. The we can use indirect inference as follows: 1. Estimate (possibly misspecified) VAR from data. A VAR is simply OLS on: Yt = π1 Yt−1 + ... + πp Yt−p + ut where Yt is the vector of observed variables at time t. In this example, Yt might be ct and it . 2. Given β, simulate model, estimate VAR from simulations, repeat until minimize objective function Consistency A general theorem on consistency is: Theorem 1. If (i) Q(θ) is uniquely minimized at the true parameter value θ0 , (ii) Θ is compact, (iii) Q(·) p p is continuous, and (iv) supθ∈Θ |Q̂(θ) − Q(θ)| → 0, then θ̂ → θ0 . We will now discuss applying this theorem to minimum distance. To do that, we need to verify each of the conditions: p p 1. (Identification): suppose fˆn (θ) → f (θ) and Ŵ → W , so that Q(θ) = f (θ)� W f (θ). As with GMM, showing that this function has a unique minimum is difficult. A local identification condition is that rank ∂f ∂θ = p, where θ is p × 1. Global identification is typically just assumed. 2. (Compactness): assume it. 3. (Continuity): depends on the particular application. For CMD and indirect inference, f (θ) = π − h(θ) is continuous as long as h(θ) is continuous. Since h(θ) does not depend on the data at all, this condition is easily checked. In the panel data example, h(θ) is obviously continuous. 4. (Uniform Convergence): depends on the particular application. Recall lemma 3 from lecture 2. It was: Review of Asymptotic Normality 4 p Lemma 2. Suppose Q̂(θ) → Q(θ) for each θ ∈ Θ. Then uniform convergence holds if for some h > 0, we have uniformly for θ, θ� ∈ Θ ˆ � )| ≤ BT ||θ − θ� ||h , BT = Op (1) |Q̂(θ) − Q(θ For CMD and indirect inference, ˆ (π̂ − h(θ� ))� | Q̂(θ) − Q̂(θ� ) =|(π̂ − h(θ))� Ŵ (π̂ − h(θ)) − (π̂ − h(θ� ))� W =|2(h(θ) − h(θ� )Ŵ π ˆ + h(θ)� Ŵ h(θ) − h(θ� )� Ŵ h(θ� )| ≤|2(h(θ) − h(θ� )Ŵ π̂ | + |h(θ)� Ŵ h(θ) − h(θ� )� Ŵ h(θ� )| ≤|2(h(θ) − h(θ� )Ŵ π̂| + |(h(θ) − h(θ� ))� Ŵ (h(θ) − h(θ� ))| so a sufficient condition is that h(θ) is Hölder continuous on Θ, i.e. |h(θ) − h(θ� )| ≤ K||θ − θ� ||h for some h > 0 and all θ, θ� ∈ Θ. A sufficient condition for Hölder continuity is that h() is differentiable with a bounded derivative because then |h(θ) − h(θ� )| ≤ sup ||h� ||||θ − θ� || Θ Clearly, this condition holds for the panel data example. It could also be checked in other applications. • If h(θ) is computed through simulation, then some additional steps need to be taken to show consistency. Let hS (θ) denote the value of h(θ) computed from S simulations. Typically, hS (θ) p ˆ to will be some standard estimator and we will know that hS (θ) → h(θ) as S → ∞. For Q converge uniformly, we need to promise that S → ∞ as T → ∞, and we will need the convergence of hS to h to be uniform in addition to the conditions above. Review of Asymptotic Normality Recall the basic asymptotic normality theorem from lecture 3: p Theorem 3. Asymptotic Normality If θ̂ → θ0 and (i) θ0 ∈ int(Θ) (ii) Q̂(θ) is twice continuously differentiable in a neighborhood, N , of θ0 (iii) √ d n�Q̂(θ0 ) → N (0, Ω) p (iv) There is J(θ) that is continuous at θ0 and supθ∈N ��2 Q̂(θ) − J(θ)� → 0 (v) J = J(θ0 ) is nonsingular then, √ d n(θ̂ − θ0 ) → N (0, J −1 ΩJ −1 ) Make sure that you understand the reasoning behind this result – taking a mean-value expansion of the first order condition. Asymptotic Linearity and Influence Functions 5 Asymptotic Linearity and Influence Functions Another way of describing results on asymptotic normality is by considering asymptotically linear estimators and their influence functions. θ̂ is asymptotically linear with influence funtion ψ(z) if: � √ √ n(θ̂ − θ0 ) = ψ(zi )/ n + op (1) (2) with Eψ(z) = 0 and E[ψ(z)ψ(z)� ] < ∞. Most common estimators are asymptotically linear. For example, MLE has influence function ψM LE (z) = −H −1 � ln f (z|θ0 ) We probably will not talk about asymptotic linearity or influence functions much in this course. Two places where influence functions come up are in calculating semi-parametric efficiency bounds, and in analyzing robustness (to outliers) of estimators. Asymptotic Normality of Minimum Distance In the last recitation we talked about minimum distance estimators, which have the form: θ̂ = arg min fˆn (θ)� Ŵ fˆn (θ) GMM and MLE fit into this framework, as well as classical minimum distance (CMD) and indirect inference. p CMD and indirect inference use fˆn (θ) = π̂ − h(θ) where π̂ → π0 = h(θ0 ). Let’s specialize the generic asymptotic normality theorem to minimum distance. Conditions (i)-(v) above become: p Theorem 4. Asymptotic Normality for Minimum Distance If θ̂ → θ0 and (i) θ0 ∈ int(Θ) (ii) fˆn (θ) is continuously differentiable in a neighborhood, N , of θ0 (iii) √ ˆ d nfn (θ0 ) → N (0, Ω) • For CMD and indirect inference √ ˆ √ nf (θ0 ) = (π̂ − π0 ) + op (1), so it is enough that n(π̂ − d π0 ) → N (0, Ω) p (iv) There is G(θ) that is continuous at θ0 and supθ∈N ��fˆn (θ) − G(θ)� → 0 • For CMD and indirect inference, �fˆn (θ) = �h(θ), so it is enough that h(θ) is continuously differentiable. p (v) Ŵ → W is positive semi-definite and G� W G is nonsingular then, √ d n(θ̂ − θ0 ) → N (0, (G� W G)−1 G� W ΩW G(G� W G)−1 ) Proof. Verify that these conditions are the same as in theorem (3). The primary difference compared to the basic asymptotic normality theorem is that twice differentiability of the objective function is restated as once differentiability of the distance function. Example 5. Chamberlain Panel Data: See recitation 1 notes for setup. π are unrestricted OLS coefficients. h(θ) = IT ⊗ β � + ıT λ� . We know that OLS is asymptotically normal, so condition (iii) is satisfied. h() is linear, so conditions (ii) and (iv) hold. Suppose we choose Ŵ = Ω̂−1 , where Ω is the usual OLS estimate of p ˆ → Ω is positive semi-definite. G� W G will be nonsingular as long the variance of π. We know that Omega as the model is identified. If there are K regressors, and T time periods, then there are T 2 K elements in π. There are K unknowns in β and T K unknowns in λ. Hence, an order condition is that T ≥ 2. Variance Matrix Estimation 6 Variance Matrix Estimation To use results on asymptotic normality for inference, we need to be able to consistently estimate the asymp­ totic variance matrix. The Hessian term, H, for MLE and Jacobian, G, for GMM can simply be estimated by evaluating the derivative of the sample objective function at θ̂. Estimation of the middle term, the variance � of the gradient, depends � on whether there is dependence in the data. For iid data, Ω = E[�qˆi (θ0 )�qˆi (θ0 ) ], which when Q(θ) = qi (θ), can be estimated by � ˆ= 1 Ω �q̂i (θ̂)�q̂i (θ̂)� n p Lemma 4.3 from Newey and McFadden gives precise conditions for when Ω̂ → Ω Lemma 6. Newey and McFadden Lemma 4.3 If zi is iid, a(z, θ) is continuous at θ0 and there is a neighbor­ � � � p p ˜ → hood, N , of θ0 , such that E supθ∈θ0 �a(z, θ)� < ∞, then for any θ̃ → θ0 , we have n1 a(zi , θ) E[a(z, θ0 )]. When the data is not iid, Ω = � E[�qˆi (θ0 )�qˆi (θ0 )� ], and some other estimator must be used. The same ideas that apply to OLS apply here. For example, if there is clustering, then � 1 �� ˆ= 1 Ω �q̂i (θ̂)�q̂j (θ̂)� C c nc i j is a consistent estimator for Ω. If there is serial correlation, then Newey-West or some similar estimator can be used. You can learn more about this time series if you want. GMM The above remarks apply to GMM with g(zi , θ) in place of �q̂i (θ). MLE For MLE, we know that √ d n(θ̂ − θ0 ) → N (0, H −1 ΩH −1 ) where H = E[�2 ln f (z |θ)] and Ω = E[� ln f (z |θ)� ln f (z |θ)� ]. In lecture 4, we saw that when the likelihood is correctly specified the information equality holds, Ω = H −1 . This suggests the following estimators for the asymptotic variance: � � 2 �−1 ∂ ln fi • Hessian: Ĥ −1 = n1 | ∂θ∂θ � θ=θˆ – In principle, when doing conditional MLE you can also use the expected conditional hessian: � � 1� ∂ 2 ln f (yi |xi , θ̂) ĤE = E |xi ∂θ∂θ� n but it is often difficult to compute this expectation ˆ ln f (z|θ) ˆ� ˆ a = 1 � � ln f (z|θ)� • Outer product of gradients: Omeg n • Sandwich: Ĥ −1 Ω̂Ĥ −1 – Could use ĤE in place of Ĥ – Since this estimator does not use the information equality, it is consistent even if the likelihood is misspecified (as long as θ̂ remains consistent) Hypothesis Testing 7 Hypothesis Testing Suppose we want to test a hypothesis of the form: H0 : r(θ) = 0 where r : Rk → Rq is differentiable. First we will discuss the familiar likelihood setup, then we will talk about testing in GMM. Before discussing these test statistics, it will be useful to review the delta method and to derive the asymptotic distribution of a constrained extremum estimator. Delta Method Suppose d � � √ d n(θ̂ − θ0 ) → N (0, V ). Let f (θ) be continuously differentiable. Then √ n(f (θ̂) − � f (θ0 )) → N (0, f (θ0 )V f (θ0 ) ). Asymptotic Normality of Constrained Estimators Suppose θ̂ solves: θ̂ = arg min Q(θ) s.t. r(θ) = 0 The first order condition for this problem is: � 0= � �Q(θ̂R ) + λr� (θ̂R ) r(θ̂R ) Expanding around θ0 and λ0 = 0 gives: � � � �� 2 � �Q(θ0 ) + λ0 r� (θ0 ) θ̂ − θ0 � Q(θ̄) r� (θ̄)� 0= + R r� (θ̄) r(θ0 ) r(θ0 ) λ̂ − λ0 � � � 2 � � � −1 � � √ � Q(θ̄) r (θ̄) �Q(θ0 ) θˆ − θ0 n R = � r (θ̄) 0 0 λ̂ � 2 −1 �� � √ (� Q) − (�2 Q)−1 R̄� (R̄(�2 Q)−1 R̄� )−1 R̄(�2 Q)−1 (�2 Q)−1 R̄� (R̄(�2 Q)−1 R̄� )−1 �Q(θ0 ) = n −(R̄(�2 Q)−1 R̄� )−1 0 (R̄(�2 Q)−1 R̄� )−1 R̄(�2 Q)−1 � � �� 2 −1 √ 2 −1 � 2 −1 � −1 2 −1 ( n�Q(θ0 )) (� Q) − (� Q) R̄ (R̄(� Q) R̄ ) R̄(� √ Q) = (R̄(�2 Q)−1 R̄� )−1 R̄(�2 Q)−1 ( n�Q(θ0 )) ¯ = r� (θ̄). This gives us the following conclusions: where R Theorem 7. Under the conditions of theorem 3 andj r(θ) is continuously differentiable in a neighborhood of θ0 , we have: � � √ √ n(θˆR − θ0 ) = J −1 − J −1 R� (RJ −1 R� )−1 RJ −1 ( n�Q(θ0 )) + op (1) (3) √ √ −1 � −1 −1 ˆ =(RJ R ) RJ ( n�Q(θ0 )) + op (1) nλ (4) √ √ −1 n(θ̂ − θ0 ) =J ( n�Q(θ0 )) + op (1) (5) √ √ −1 � −1 � −1 −1 ˆ n(θ̂ − θR ) =J R (RJ R ) RJ ( n�Q(θ0 )) + op (1) (6) √ √ � −1 � −1 −1 n�Q(θ̂R ) = − R (RJ R ) RJ ( n�Q(θ0 )) + op (1) (7) Proof. (3) and (4) are direct consequences of the previous reasoning. (5) comes from theorem 3. (6) is simply the difference of (3) and (5). (7) comes from pluggin (4) into the first order condition. This theorem tells us the asymptotic variance of various quantities that will be used in our test statistics. For example from (6), we know that � � √ d n(θˆ − θ̂R ) → N 0, J −1 R� (RJ −1 R� )−1 RJ −1 ΩJ −1 R� (RJ −1 R� )−1 RJ −1 ML Testing 8 ML Testing When doing MLE, we have the usual trinity of tests: Wald, Lagrange multiplier, and likelihood-ratio. Throughout we will write Avar(θ̂) to denote the asymptotic variance of θ̂. It should be straightforward to calculate using theorem 7. Each statistic can be thought of as a measure of the difference between the restricted and unrestricted objective functions. The Likelihood ratio uses the actual difference. The Wald statistic uses a quadratic expansion at the unrestricted estimate, θ̂, to approximate the restricted objective value. The Lagrange multiplier uses a quadratic expansion at the restricted estimate, θ̂R , to approximate the unrestricted objective value. Wald Wald test statistics look at the distance between θ or r(θ) in the restricted and unrestricted models. One version of the Wald statistic is motivated by asking whether r(θ̂) = 0? It uses the test statistic: W1 =nr(θ̂)� AV ar(r(θ̂))−1 r(θ̂) (delta method) (8) d =nr(θ̂)� (r� (θ̂)AV ar(θ̂)r� (θ̂))−1 r(θ̂) → χ2q (9) Another variant of the Wald test looks at the distance between restricted and unrestricted estimates of θ: d W2 =n(θ̂ − θ̂R )� (Avar(θ̂ − θ̂R ))−1 (θ̂ − θ̂R ) → Xq2 (10) Lagrange Multiplier The Lagrange multiplier test is based on the fact that under H0 , the Lagrange multiplier of the restricted optimization problem should be near 0. The first order condition from the restricted ML is: 1� ˆ � (θˆR ) � ln f (z|θ̂R ) = λr n which suggests the test statistic: LM1 = �� �� � 1 �� d � ln f (z|θ̂R ) Avar(� ln f (z|θ̂R ))−1 � ln f (z|θ̂R ) → χ2q n Equivalently, we could look at the estimated Lagrange Multiplier, LM2 = nλ̂� Avar(λ̂)−1 λ̂ Likelihood Ratio The Likelihood ratio statistic compares the restricted and unrestricted likelihoods. d LR = 2(LN (θ̂) − LN (θ̂R )) → χ2q To prove this, expand LN (θ̂R ) around LN (θ̂): LR =2(LN (θ̂) − LN (θ̂R )) � � =2 LN (θ̂) − LN (θ̂) − �LN (θ̂)(θ̂ − θ̂R ) − (θ̂ − θ̂R )� �2 LN (θ̄)(θ̂ − θ̂R ) =(θ̂ − θ̂R )� �2 LN (θ̄)(θ̂ − θ̂R ) GMM 9 GMM The same three test types of test stastistics work for GMM. The Wald and Lagrange Multiplier statistics are particularly identical to the ML case. The likelihood ratio statistic is replaced by the distance metric: DM = 2n(Qn (θ̂) − Qn (θ̂R )) Asymptotics of Simulated Estimators This is a quick and dirty discussion of the asymptotics of simulated extremum estimators. See Kenneth Train’s book and the references therein for more details and rigor. Train’s book can be found at http: //elsa.berkeley.edu/books/choice2.html. Let Qn (θ) denote the exact objective function. Let Q̃n (θ) denote the simulated objective function. Assume that θ̂ = arg min Qn (θ) is consistent and asymptotically normal. We want to understand the behavior of θ̃ = arg min Q̃n (θ). For specificity, assume that in simulating, we make R draws for each observation, and these draws are independent across observations. Consistency For consistency, the key condition to check is that Q̃(θ) = plim Q̃n (θ) is uniquely minimized at θ0 . Consider the first order condition: � � � � �Q̃n (θ) =�Qn (θ) + Er (�Q̃n (θ)) − �Qn (θ) + �Q̃n (θ) − Er (�Q̃n (θ)) where Er denotes an expectation taken over our simulated draws. If we can show that the second and third terms on the right vanish as n → ∞, then we will have consistency. The third term is easy. Since we are making R independent draws for each observation, as long as R is fixed or increasing with N , �Q̃n (θ) satisfies an LLN and converges to its expectation. The second term depends on how we are simulating. If R increases with N , then it also vanishes because of an LLN. Furthermore, even if R is fixed with N , it will be zero, if our simulations result in an unbiased estimate of the gradient. In the mixed logit example above, the simulation of choice probabilities is unbiased. Therefore, NLLS, for which the first order condition is linear in P̃ , is consistent with fixed R. However, MLE, for which the first order condition involves P1˜ , is consistent only if R increases with N . For this reason, people sometimes suggest using the method of simulated scores (MSS) instead of MSL. MSS call for simulated the score in an unbiased way and doing GMM on the simulated score. Asymptotic Normality As always, we start by taking an expansion of the first order condition: √ √ n(θ̃ − θ0 ) =(�2 Q̃n (θ̄))−1 ( n�Q̃n (θ0 )) p If θ̃ is consistent, then (�2 Q̃n (θ̄))−1 → (�2 EQ̃(θ0 ))−1 . The main thing to worry about is the behavior of the gradient. As above, it helps to break it into three pieces: � √ � � √ √ √ � n�Q̃n (θ0 ) = n�Qn (θ) + n Er (�Q̃n (θ)) − �Qn (θ) + n �Q̃n (θ) − Er (�Q̃n (θ)) Let’s start with the third term. Suppose we have iid observations so that �Q̃n = variance of �q̃i,1 . Then the variance of �q̃i,R is S/R, and � √ � d n �Q̃n (θ) − Er (�Q̃n (θ)) → N (0, S/R) �n i=1 �q̃i,R . Let S be the Asymptotic Normality 10 Now, on to the second term. As above, it is zero if our simulations are unbiased. If our simulations are √ biased, then it is O( R1 ). If R is fixed, then our estimator is inconsistent. If Rn → 0, then this term vanishes, and our estimator has the same √ asymptotic distribution as when using the exact objective function. If R grows with n, but slower that n, then θ̃ is consistent, but not asymptotically normal.