20
Standard tests Identification and Overidentification Testing Overidentifying Restrictions IV/GMM Inference. Identification and Overidentification Walter Sosa-Escudero Econ 507. Econometric Analysis. Spring 2009 March 8, 2009 Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

Identification

Embed Size (px)

DESCRIPTION

Econometria

Citation preview

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    IV/GMM Inference. Identification andOveridentification

    Walter Sosa-Escudero

    Econ 507. Econometric Analysis. Spring 2009

    March 8, 2009

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    Standard tests

    Standard inference comes directly from the asymptotic normality result

    n (g 0) d N(0,AV(g))

    Then, it is easy to see that

    Under H0 : k = k0

    tk =ngk k0

    AV(gk)

    p N(0, 1)

    where AV(gk) is any consistent estimator.

    Under H0 : R0 r = 0

    W = n(Rg r)[RAV(g)R]1(Rg r) p 2(q)with R, r and q defined as in our original large sample inferenceproblem.

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    Overidentification: Prelude

    Suppose there are two instruments z1 and z2, for a modely = x0 + u where x is a single explanatory variable.Suppose invalid instrument means that for no E(zijui) = 0, j = 1, 2.

    You use only one instrument, say z1 and get ui = yi xiwhere is the IV estimator using z1 as instrument.

    What can you learn about instrument validity from thecorrelation between ui and z2i?.

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    Identification and Overidentification

    Recall that our moment conditions for the IV case are

    E[zi(yi xi0)] = 0,

    and our problem was that the sample counterpart

    1n

    ni=1

    zi(yi xib) = 0

    implies a system of p linear equations with K unknowns, whichcannot produce a solution when p > K.

    If all the moment conditions are valid we could discard pKmoment conditions in any arbitrary way and produce a MMestimator using the K moment conditions retained.

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    Another thing we can do is to combine the p moment conditionslinearly so as to produce K linearly independent momentconditions, that is, use:

    BE[z(yi xi0)] = 0

    where B is any K q matrix with (B) = KIn fact throwing away moment conditions implies a particularchoice of B! (which one?).

    Note that the sample counterpart

    1nB

    ni=1

    zi(yi xib) = 0

    produces a single solution (why?).

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    GMM as a particular case of MM?

    GMM is choosing a particular B. Let ui(b) yi xib andu(b) Y Xb. Recall that the GMM estimator minimizes:

    J(b) = n{1nu(b)Z

    }Wn

    {1nZ u(b)

    }so the FOCs are:

    X Zn

    WnZ u(b)n

    = 0

    which can be seen as the sample counterpart of:

    E[xizi]W Kp

    E[ziui(0)] p1

    = 0

    which are K moment conditions!

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    E[xizi]W Kp

    E[ziui(0)] p1

    = 0

    Then

    GMM is choosing a particular way of linearly combining the pmoment conditions so the system becomes exactly identified.

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    Identifying and Overidentifying Restrictions

    It is very illuminating to explore how these linear combinations areproduced.

    First note that W can be written as W =W 1/2W 1/2, where

    W 1/2 is an invertible p p matrix.It will be more convenient to re-express our original p momentconditions as follows:

    W 1/2E[ziui(0)] = 0

    Note that this does not alter at all the informational structure ofthe problem (why?).

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    Now write the GMM version of the moment conditions as:

    E[xizi]W E[ziui(0)] = 0

    E[xizi]W1/2W 1/2 E[ziui(0)] = 0

    F W 1/2E[ziui(0)] = 0

    where F E[xizi]W 1/2 . Note F is K p with rank K.Let h W 1/2E[ziui(0)], so the moment conditions imply h = 0.

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    h =W 1/2E[ziui(0)] is a p vector, and it can be decomposedorthogonally as

    h = Ph+Mh

    where P and M are any pair of orthogonal projection matricesthat project h orthogonally onto some subspace and its orthogonalcomplement.

    Note that the moment condition implies Ph = 0 and Mh = 0.

    Also, note that the decomposition holds for any pair of projectionmatrices.

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    In particular for P = F (F F )1F Pf , whereF = E[xizi]W

    1/2 .

    Note F is K p with rank K, hence Pf is the projection matrixthat projects vectors of dimension p onto the span of the Kcolumn vectors in F , hence Pf has rank K andMf = I F (F F )1F has rank pK

    (Remember the dimension theorem...).

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    ThenF (F F )1F h = 0,

    forms a system of K linearly independent equations with Kunkowns.

    Now note since F (F F )1F has rank K, F (F F )1F h = 0whenever

    F h = 0

    which are the MM conditions derived from the GMM procedure.

    This leds us to a very important result.

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    What GMM is doing in when p > K is decomposing themoment conditions, h, in two orthogonal parts. One that isused to exactly identify the relevant parameters (Pfh = 0)and another part which is left unused (Mfh = 0).The moment conditions Pfh = 0 are called the identifyingconditions and the conditions Mfh = 0 are called theoveridentifying conditions.

    In a certain sense, we can see GMM as using a particular setof K moment conditions and discarding the remaining pK.

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    Another interesting intuition is the following.

    Let h W 1/2n n1n

    i=1 ziui(g). Then the minimized GMMobjective function can be expressed as:

    J(g) = n hh

    Let Fn n1 X Z W 1/2

    n , Pfn = Fn(F nFn)1F n andMfn = Ip Pfn.By orthogonal decomposition, h = (Pfn +Mfn)h. Replacingabove and using the properties of these matrices:

    J(g) = n[h(Pfn +Mfn)(Pfn +Mfn)h

    ]= n

    [h(Pfn +Mfn)h

    ]= n

    [hPfnh+ hMfnh

    ]Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    J(g) = n[hPfnh+ hMfnh

    ]The FOC of the GMM problem sets Pfnh = 0, which leds to avery useful intuituion about how GMM works and its usefulness fortesting:

    The GMM estimator satisfies strictly the identifyingrestrictions and tries to make the overidentifying restrictionsas small as possible.

    The minimized value of J(g) = n hMfnh measures how faris the sample from satisfying the overidentifying restrictions.

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    Testing overidentifying restrictions

    Intuituion:

    The fact that J(g) measures how far is the sample fromsatisfying the overidentifying restrictions can be exploited todesign a formal specification test.

    If all the assumptions of the overidentified GMM model hold,then asymptotically the overidentifying conditions should besatisfied: the identifying conditions succeed in producing aconsistent estimator and hence force all moment conditions tohold.

    If the J(g) is too large then some of the conditions thatguarantee consistency and asymptotic normality are likely tobe false (more on this later).

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    The J test: the test statistic for the overidentifying restrictionstest is:

    J = Jn(g) = nu(n)Z

    nS1n

    Z u(i)n

    and under H0 : E[ziui(0)] = 0 it converges in distribution to2(pK).

    First note that the test uses the optimal GMM estimator, thatis, Wn = S1n .Intuition: the J test checks if the GMM is small. Accordingto our previous result, this means checking if theoveridentifying restrictions are small.

    Under the null that the model is correctly specified (all GMMassumptions hold), GMM is consistent and hence theoveridentifying restrictions should be close to zero.

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    What happens under the alternative hypothesis?

    Suppose the rank condition holds. The sample version of theidentifying conditions must be satisfied, so g

    p +.Asymptotic normality of the moment conditions also holds,but centered at some other place that is not zero.

    Rewrite the statistic as:

    J = nu(n)Z

    nS1n

    Z u(i)n

    Then when H0 does not hold, n and the remaining partconverges to something not centered at zero, then J .It is a global test for misspecification. It may mean that somemoment conditions are invalid and/or that the model ismisspecified.

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    Proof: By our previous result

    Jn(g) = n[hMfnh

    ]with h W 1/2n Z

    u(g)n , so

    Jn =

    {W 1/2n

    Z u(g)n

    }Mfn

    {W 1/2n

    Z u(g)n

    }From the asymptotic normality proof of GMM we got

    Z u(g)n

    d N(0, S).

    If Wn = S1n and W1/2n = S1/2

    W 1/2nZ u(g)

    n= S1/2

    Z u(g)n

    d N(O, Ip)

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

  • Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions

    Hence, the J test is an quadratic form of a p vector of independentnormal variables, normed by an idempotent matrix with rankpK, then the result follows.

    Recall that if y N(0, Ip) and A is an idempotent matrix with rank q, then yAy 2(q) (Hayashi (2000,p. 37)).

    Walter Sosa-Escudero IV/GMM Inference. Identification and Overidentification

    Standard testsIdentification and OveridentificationTesting Overidentifying Restrictions