Digitized by the Internet Archive in 2011 with funding from Boston Library Consortium IVIember Libraries http://www.archive.org/details/inferenceonparam00cher2 i^tlW^Y HB31 .M415 Massachusetts Institute of Technology Department of Economics Working Paper Series INFERENCE ON PARAMETER SETS IN ECONOMETRIC MODELS & Confidence Regions for Parameter Sets Econometric Models," id# 963451) (See recent revision called "Estimation in Chemozhukov Han Hong Elie Tamer Victor Working Paper 06-1 8A June 16,2004 Room E52-251 50 Memorial Drive Cambridge, MA 02142 This paper can be downloaded without charge from the Social Science Research Networl< Paper Collection at http;//ssrn.com/abstract=909625 ^ '•^'^^^^^i^f^LSI^'™''! FEB 2 2 2007 "libraries INFERENCE ON PARAMETER SETS IN ECONOMETRIC MODELS* VICTOR CHERNOZHUKOVf HAN HONG^ ELIE TAMER* Abstract. This paper provides confidence regions for minima of an econometric criterion function Q{d). The minima form a set of parameters, Q/, called the identified set. In economic applications, 0/ represents a class of economic models that are consistent with the data. Our inference procedures are criterion function based and so our confidence regions, which cover 0/ with a prespecified probability, are appropriate level sets of Qn{S), the sample analog of Q{9). When 0/ is a singleton, our confidence sets reduce to the conventional confidence regions based on inverting the likelihood or We show that our procedure is valid under general yet simple conditions, and we provide feasible resampling procedure for implementing the approach in practice. We then show that these general conditions hold in a wide class of parametric econometric models. In order to verify the conditions, we develop methods of analyzing the asymptotic behavior of econometric criterion functions under set identification and also characterize the rates of convergence of the confidence regions to the identified set. We apply our methods to regressions with interval data and set identified method of moments problems. We illustrate our methods in an empirical Monte Carlo study based on Current Population Survey data. other criterion functions. Key Words: Set estimator, level sets, interval regression, subsampling bootsrap Date: Original Version: May 2002. This * Note: This paper and rates is currently undergoing 2004 version: June, MAJOR . revision following referee of convergence are added. Convexity assumptions comments. New results on estimation on the parameter space are being replaced by more general conditions. The most *We thank Alberto Abadie, Gary Chamberlain, Jiuyong Hahn, Bo Honore, Sergei Izmalkov, Guido Imbens, Christian recent, yet preliminary and incomplete, revision can be requested via e-mail. September, 2005. Hansen, Yuichi Kitamura, Chuck Manski, Whitney Newey, and seminar participants at Berkeley, Brown, Chicago, Harvard-MIT, Montreal, Northwestern, Penn, Princeton, UCSD, Wisconsin, and participants of Econometric Society in San-Diego for ' Department of Economics, MIT, vchernSmit 0241810) § is Winter Meetings . edu. Research support from the National Science Foundation (SES- gratefully acknowledged. Department of Economics, Duke University, han.hong8duke.edu. Research support from the National Science Foun- dation (SES-0335113) J at the comments. Department of is gratefully acknowledged. Economics, Princeton University, tamerSprinceton.edu. Research support from the National Science Foundation (SES-0112311) is gratefully acknowledged. 1 Introduction and Motivation 1. Parameters of interest econometric models can be defined as in minimize a population objective or criterion function. at a particular If this criterion tJiose parameter vectors that function parameter vector, then one can obtain valid confidence regions is minimized uniquely (or intervals) for this parameter using a sample analog of this function. Likelihood and method of moments procedures are two commonly used and well studied methods in this setting. inference to econometric models that are set identified, minimized on a identified set set of parameters, the identified set. i.e., Our This paper extends this criterion-based models where the objective function goal is to make is inferences directly on the and to provide a method of obtaining confidence regions with good properties (such as consistency and equivariance) that cover the identified set with a prespecified probability. Our point of departure analog Qn{0) where 9 e where every 9 paper this is in is @ a nonnegative population criterion function Q{0) and czW^. The identified 9/ indexes an economic model that lim for a prespecified confidence level Qn a e for can be defined as 0/ A 9/ from the level set {6> consistent with the data. is ^(9/ C C„) = (0, 1). = level sets of e 9 The sample Q{e) : = 0} objective of Qn{9) such that a, Cn{c) of the finite sample criterion function defined as Cn{c) := iO for C„ to construct confidence sets (1.1) is set its finite : Qn{d) ~ Qn some appropriate normalization Cn = Cn{c) with the cut-off level c < a„. — Ca, c/a-n \, where g„ := inf Qn{0) or g„ := 0, In order to obtain the correct coverage (1.1), we choose where c^ equals the asymptotic a-quantile of the coverage statistic: Cn := sup an{Qn{d) which is a quasi-likelihood-ratio type quantity. properties, such as consistency situations, where the and equivariance identified set is - g„), The constructed confidence to reparameterization. defined as the minimand sets possess Our approach important covers general of an objective function. In addition, our confidence regions are sets that are robust to the failure of point-identifying assumptions in that they cover the unknown identified set - whether These confidence where the We it is a set or a point - with a prespecified probability. sets collapse to the usual confidence intervals identified set 9/ is show that the above based on the likelihood ratio in cases a singleton. level set Cn[ca) has correct asymptotic coverage, where c^ is the ap- propriate quantile of a well defined coverage statistic C, the nondegenerate large sample limit of Cn- Then, we provide general resampling methods to consistently estimate the percentile Cq. The coverage and resampling results hold under general (high level) conditions. econometric models, we then show that these conditions are satisfied. In Focusing on a class of the process, acterizations of the stochastic properties of the criterion functions process, an{Qn{-) the fact that the population criterion Q is minimized on a set rather we provide — g„), char- exploiting than a point. Furthermore, we obtain the asymptotics of the coverage statistic C„ and of the level sets Cn{c), focusing on coverage and speed of convergence of Cn{c) to 9/. The paper then considers two important applications: We regression with interval censored outcomes and set-identified generalized method-of-moments. finally illustrate our methods in a Monte Carlo study based on data from the Current Population Survey. In the last decade, a growing body of literature has considered the problem of inference in partially identified models, (2003). i.e., models where parameters of interest are While most of the work, e.g. is paper is cf. Manski Horowitz and Manski (1998) and Imbens and Manski (2004), exclusively focuses on the case where there this set identified, is a scalar parameter of interest that lies in an interval, concerned with inference on vector parameters in problems where the identified set defined via a general optimization problem, as in the economic problems described below. multivariate set is This usually not an interval (or cube) and can be a set of isolated points or manifolds. Moreover, the methods used in the interval case are based on estimating the two end-points of the interval. Hence they are not applicable in the general set-identified case, for which a different approach is needed. The main motivation examples. for posing the identified set as the object of inference In Hansen, Heaton, and Luttmer (1995), the identified set models that obey the pricing-error and models with multiple is equilibria supported across markets or industries. (2003). Another example is E through a birth).-' Z flexible, is There is no single "true" equilibrium that is example where potential income Y= ao + aiE -\- a-iE'^ Y + is related to e. Although not point-identified in the presence of the standard quarter- suggested in Aiigrist and Krueger (1992) (indicator of the In the absence of point identification, some In the set of parameters that describe different quadratic functional form, the instrumental orthogonality restriction In market returns. vary across observational units; see Ciliberto and Tamer in the following parsimonious, this simple model of-birth instrument many the structural instrumental variable estimation of returns to schooling. Suppose that we are interested education may motivated by a subset of asset-pricing volatility constraints implicit in asset equilibria, the identified set played, since particular equilibria is is all first quarter of of the parameter values (qqi 0^11^2) consistent with E{Y — ao + Qi-E -f a2£^'^)(l, Z^ = are of interest for cases the indicators of other quarters of birth are used as instruments. However, these instruments are not correlated with education (correlation is extremely small) and thus bring no additional identification information. 3 moment and purposes of economic analysis. Similar partial identification problems arise in nonlinear instrumental variables problems, see e.g. Literature. To our knowledge, the models can be found in the Demidenko (2000) and Chernozhukov and Hansen earliest work in econometrics paper of Marschak and Andrews (1944). on parametric (2001). set identified There, the identified set is the collection of parameters representing different production functions that are consistent with the data and functional restrictions the authors consider.^ Gisltein and Leamer (1983) provide set consistent estimation in a class of likelihood models where 0/ as the set of parameters that are robust to misspecification. Klepper and Leamer (1984) generalize the Frish bounds to multivariate measurement regression models with errors. On the other hand, Hansen, Heaton, and Luttmer (1995) provide consistent set estimates of means and standard deviations in a class of asset pricing methods. Manski and Tamer (2002) provided conditions under which an appropriately defined estimates the identified about the identified set. However, these consistency results do not contain a method set. To the best of the paper is for inference of our knowledge, there are no results in the literature that deal with the general problem of obtaining confidence regions The remainder set consistently for organized as follows. parameter sets.^ Section 2 provides a general theory of inference on the identified sets based on general criterion functions. Section 3 focuses on a class of regular parametric models, provides verification of the regularity conditions posed in Section provides additional results that pertain to regular cases. Section 4 provides a of the methods, and Section 2. forms the basis General Set Inference for the rest of the analysis. We first in Large Samples this section The Wn), where data {Wi, criterion function 0/, the identified which 0/ and formalize some ..., W„} are defined is based on a criterion function Qn{S) on some common probabihty space Qn{0) converges to a continuous criterion function Q{9), that (Basic Setup). Criterions (5„ K*^ : — M+ » and For a good description of Marschak and Andrews (1944), see Chapter The only other paper known to us is that can be estimated.) The problem is interval-identified of inference Imbens and Manski (2004) and Q III of : E'^ — > E"*" and is (fi, = F, P). minimized at in about 9' Appendix G. 4 is (i.e. satisfy Nerlove (1965). Imbens and Manski (2004). Their paper considers a inference about the a real parameter 6' that in result set. Assumption A.l shown we present our main define the identified set For given data, the inference about the parameter set 0/ ..., Monte Carlo evaluation be used throughout. definitions that will Qn{d, Wi, and 5 concludes. Generic Inference on Identified Sets. In 2.1. 2, different problem of contained between some upper and lower bounds fundamentally different from inference about G/, as e i. is Q{9) ii. continuous and Qn{d) is = iii. 0/ iv. QiOj) = V. Qn{0) - Q{B) = argminege[Q(0)] a is when 0/ hyperplanes) is eQ. states a standard compactness finite It also as the minimizer of the limit criterion function Q. 0/ defines and convexity assumption, which are important union of compact connected a finite sets, is a an assumption that serves to organize union of isolated points and the finite union of compact sets with boundaries defined by manifolds (nonlinear The assumption . a finite union of connected compact subsets ofQ, is This covers both the case when 0/ the presentation. case lower-semicontinuous, is Op(l) for each 9 to the subsequent analysis. region 0/ BE RELAXED], to for each Oj £ 0/, Assumption A.l The [CONVEXITY o compact and convex subset ofR'' that Q{di) Q convergence condition serves to relate = The pointwise serves as a convenient normalization. as the limit of the finite sample objective function. The pointwise convergence will be strengthened later on. Lemma 2.1 shows how to construct a level set of the sample objective function that The provide the proper inferential statement (1.1) about 0/ in a generic setting. objective function Q„ is later. is Note that choosing g„ Example e.g. -.an (Q„((9) - g„) < c[, where g„ := inf Q„(6') or g„ := 0, defined in Assumption A. 2 below. non-empty, though 1 in c-level set of given by Cnic) := \e where a„ will eventually = Typically a„ equals n in the regular cases studied infeeeQn(^) guarantees that the confidence region Cn{c) some cases one has inf^ge Qn{9) = is always with probability converging to one, see in Section 3. Consider the following coverage index fo P{c)= 1") ^ c - snp an {Qn{9) - The 0/ C Cn(c) sign of the index p(c) indicates whether such that 9 ^ Cn{c), we have a„ {Qn{9) index is also linear in property is c, which — > q-n) will allow us to c or not. For example, which implies that p{c) have data-dependent < 0, if there and vice cut-off' levels c. £ 0/ is versa. The Another desirable invariance of the index p to parameter transformations which implies equivariance of Cn{c) For instance, a one-to-one transformation of parameters 9 to parameter transformations. changes the level set in the equivariant {t{9) The Qn) eee, ' following : a„ {Qn{T-\T{e))) lemma summarizes For definition of manifold, see e.g. way the discussion. Milnor (1964) . -qn)<c}=T {Cn{c)) — » t{9) Lemma p{c) Assumptions A.l-(ii) and 2.1. Let ^ < 0/ 2 Cn{c) cut-off level; and p the coverage index III. > (c) Then, A.l-(iii) hold. •^ 9/ C Cn{c); II. the coverage property holds: I. the coverage index is linear in the invariant to re-parameterization; and IV. the level sets are is equivariant to bijective parameter transformations. Next we state the Assumption A. 2 main condition that enables (Coverage large sample inference Suppose that there Statistic). on 0/. sequence of constants a„ exist a ^ oo such that Cn= where C is a nondegenerate we explain how In the sequel interest, random where a„ = n. We sup an{Qn{Sl)- Qn) -^d C, variable. this main assumption methods also provide new asymptotic methods Assumption A. 2 leads us attained in suiRciently regular models of for verification of this the limit variable C. Verification of Assumption A. 2 set of is is a difficult assumption and for finding matter and requires developing a of dealing with criterion functions under set identification. main theorem, which provides a generic to the following on result inference in set-identified models. Theorem 2.1 (Generic Inference on Sets). Suppose Assumptions A.l continuity point of distribution function of C such that p{Cn{ca)) I. The result follows Op(l) -^d C — Ca- P{ei C —»d Ca- immediately from C and lim II. Lemma 2.1 P{C < and A. 2. C„(Ca)} = P[p{Cn{Ca)) Then for any Cq -^p P{ 0/ C C„(Cq) First, p{Cn{c)) \ = = a Cq,, a. €„ > O} = P{ca - C„ > = -P{C„ < c„ + Op(l)} = P{c < some criterion function. to a singleton {9i} so that the coverage statistic = a. is — Ca = Cn — Ca + O} c„} + o(l), C. from Theorem 2.1 that our method of constructing valid confidence sets builds on the classical principle of inverting Cn = that Ca Second, provided Cq a continuity point of distribution function of It is clear Cq} and A. 2 hold and On [Qni^l) — Qn) , which foUows well known involved quantity, being the supremum Indeed, in point identified cases 0/ reduces becomes a standard likelihood ratio type quantity limit laws. In the general case, the statistic over the elements of 0/: C„ = supg^^Q^ a„ {Qni^l) is a more — qn) In practice, our procedure for constructing the confidence regions critically depends on being able to consistently estimate Cq, the a-quantile of the limit variable C. In 6 all examples that we study, C non-standard and is we distribution depends on 0/. Despite this problem, its show how will to obtain a consistent estiniate of Cq by the following method. C„ We Feasible Inference. 2.2. = sup^gQ a„ {Qn{0) — first Since Qn) 6/ = (2.2) where A; we can is need to obtain an approximation to the sampling distribution of we do not observe 9/, we where C„(fc), G fc [ci,C2] • wp — Inn = 0.) The result proven initial estimate' first class of our examples, it 1, > a possibly data-dependent starting value of the cut-off. (In the use th starting cut-off k by an will replace below suggests that the asymptotic vahdity of the procedure will not depend on the starting value. In finite samples, the choice of the starting value may be important, and we discuss in Section 4. it Consider the following subsampling algorithm: For cases 1. cases when data {Wi} when {Wt} form a of the form {Wi, are construct iid, For each j = ...,Wi+i,-i}. 1, ..., (Bn) subsets of size stationary time series, construct 5„ = n — 6 ^ n of the data. b + I For subsets of size b one can use a smaller number B„ of randomly (In practice, B„ — chosen subsets under the condition that 2. all » oo as n — > oo.) Bn, compute Cj,b,n = sup - ab {Qj,biO) qj,b) eeCn{c) where Qj^ := if and Qj^ := inf^ee g„ := (3j,6((9) otherwise; Qj,b{d) denotes the criterion function defined using the j-th subset of the data only. 3. Let Cq be the a-th quantile of the sample {Cjfi^n,j 4. (Optional) As commented times using c = in Section 4.3, Ca Inn.^ (Any finite = l, , Bn}. one could repeat steps 2 and 3 number finite number of of repetitions produces the consistent estimate of Cq.)'^ We require that as n > b/n (2.3) The — choice of b oo, — > 0, The adjustment as oo for all o — > oo, and other practical aspects To guarantee asymptotic vahdity n — i?„ > factor fe — oo > at polynomial rates. of the procedure are discussed in Section 4. of the above procedure, the following assumption Inn can be replaced by In Inn or is needed. any other m{n) such that ni(n) —> oo and m{n)/n'' — » 0. The adjustment by In n is not needed in the first class of our examples. In fact, we found in Monte-Carlo experiments described in Section 4 that just two repetitions worked the best terms of coverage and computational expense. This was also confirmed by Bajari, Benkard, Levin (2003). 7 in Assumption A. 3 (Sandwich n— > oo. For k € [co,ci] 9/ C -Inn, C„(fc) C Property). Suppose that h/n lup — » — + + := {dj t < ||i|| : en,0i £ ab{ sup Qb{9) 6;", such that Cn ~* 0/} and Assumption A. 3 guarantees b — oo ai polynomial rates as > 1 is - sup Qb{0)) = Op(l), See, eeQ]" QY and a sequence of positive constants. inferential validity but also allows us to establish consistency characterize the rate of convergence of the level sets C„(A;) to the identified set 0/, as in Theorem The 2.2. know 0/, hence we intuition behind A. 3 replace it is by an estimate Cn{c). The replacement should have only a impact on the distribution of the coverage Wald it 3 Wald 9j, subsamples. We show how negligible Romano, and Wolf know (1999), p 44, in the the true 9j and replaces effect on the distribution to verify A. 3 for parametric models in Section The next theorem summarizes our 3.2). we do not similar in nature to the is hence requiring that this replacement has negligible statistic in (Theorem Politis, inference in point identified cases. There, one does not with an estimate of the subsamples. A. 3 statistic in polynomial rate of convergence assumption used by case of shown below In the subsampling bootstrap, as follows. and to results for general inference in set identified models. Theorem are iid or I. 2.2 (Consistency and Validity of Inference). Suppose that the estimation data {Wi,i form a stationary strongly -mixing sequence Then for Ca — * and that A. 2 the subsampling algorithm defined above, provided Ca, and A. 3 P{C < c} < n} = Ca, hold. continuous atc is and P(0/eC„(ca))-^a. II. The set Cn{k) for /c = Calnn is consistent under the Hausdorff metric: dH{Cn{k),Qj) < Recall that the Hausdorff metric between two sets dH{A,B) :=ms^[h{A,B),h{B,A)], rate e„ will set at the rate Cn be shown to be essentially l/^/n Under the given The conventional level of generality, is for ^> 1 0. defined as: is where Thus, under general conditions our inference method Cn{k) that conv^erge to the identified ^ en wp h{A, B) := sup mi \\a - b\\. asymptotically valid and delivers level sets with respect to the Hausdorff distance (the parametric models). subsampling appears to be the only valid resampling method. (n-out-of-n) bootstrap will not be generically consistent in the present settings. One counterexample is as follows: In Section 3, one of the leading examples, the partially identified linear IV regression, necessarily involves a parameter bootstrap fails in on the boundary problem. parameter-on-the boundary problems, we Andrews known that the (2000). Regular Parametric Case and Applications 3. In this section cf. It is establish the methods for verifying the main conditions required for implement- ing the approach, namely that existence of the limit distribution for the coverage statistic (A. 2) and the sandwich condition (A. 3). and illustrate the the We develop these methods to cover a variety of parametric examples, approach with applications to regression models with missing outcome data and generalized method of moments under we partial identification. For parametric models, establish further properties of the confidence regions, such as the speed of convergence of the level sets to the identified set, and the stochastic properties of objective functions Examples 3.1. of Parametric Problems with Set in partially identified cases. Example Identification. I. Regression with Interval-Censored Outcomes. Consider the linear conditional expectation models Ep^[Y\X] The models are not assumed = X'9, where observation on eR"^. to be correctly specified, in the sense that there £'pg[y|X] agrees with £'p[y|X] under the actual law Observed data consists 9ee,X of i.i.d. P may be no 6 such that of the data. observations (Yu, Y2i, Xi) where Yu and Y2i represents the interval Yf. Yi (3.1) e [Yu,Y2i] given Xi, a.s. In the absence of further information, the set ei = {eeR'^ (3.2) is the object of interest, as consistent with data. it : E[Yi \X] < X'9 < E[Y2\X] a.s. } represents the set of hnear conditional expectation models that are We assume that 9/ C 0, where 9 is a compact subset of R"^. Observe that 9/ minimizes the objective function (3.3) where [u)% = {uf Q{9) = > 0) x l[u Using a sample analog of (3.4) Qn{s) j [{E[Y,\x]-x'9)l + and (u)?. = {uf x l[u < {E[Y2\x]-x'0)l]dPix), 0]. Notice also that 9/ = (3.3), = IY1 {MyiA^i] - K0)l + i<n {e[Y2^\x,] - x[9y_ , {61 G : Q{9) = 0} . Manski and Tamer (2002) characterize consistent estimates no method of 0/.* However, We ence about 0/ in the sense of providing the inferential statements was given. confidence approach to inference about Example Moment Structural II. terested in deducing the set of all moment equation computed with data. The data {fl,J^,P). < {Xi,i 0/ and similar examples. in this Equations. In method moments we settings, R'^ are in- that satisfy the respect to the probability sampling distribution of the economic dj of interest are assumed to E{mi{e])] m{9, Xi) of economic models indexed by a parameter 9 £ (3.5) = provide below a n) are stationary and strongly mixing, defined on the probability space The economic models where mi{d) for infer- is = satisfy 0, a lower-semi-continuous function in 9 a.s. The entire set of models 0/ C that solve (3.5) also minimize the criterion function Q{9) (3.6) where Q{9) to nonlinear equations (3.5) Demidenko conditions E[mi{9)]'W{9)E[mi{9)], continuous for each ^ € 0, a compact subset of W^, and W{9) is positive definite matrix for each 9 e.g. = general be minimized on a where Wn{9) is In nonlinear models, the existence of multiple solutions more of a rule rather than an exception, except in special cases, see mentioned set. Qn{e) (3.7) = The may be multiple solutions as in the introduction. inference on Thus, the GMM weU 0/ may be based on the usual n[gn{e)]'Wn{e)[gn{9)], 5n(^) for inference about 0/ if the usual rank function (3.6) will in GMM function = - f^m^(e), a lower-semi-continuous, uniformly positive definite matrix.^ knowledge, no methods a continuous and Q. E. (2000). In linear models there to hold, as fail is is in the sense of providing To the best of our Neyman-Pearson confidence statements has been yet given in such settings. 3.2. General Asymptotics of Criterion Functions in Parametric Models. The prime of this section is to provide primitive conditions a wide variety of parametric models. and tools that help verify conditions A. 2 goal and A. 3 in The tools will be illustrated using the examples stated in the {5 £ 9 next section. Specifically, if £n > they show that the set 0„ = ; Qn(d) < mine Qn{0) + in} converges almost surely to the set 9/ converges to zero at a rate strictly slower than the rate at which Qti(-) converges uniformly to Q(-)- They use a Hausdorff metric as a distance function between For example, it is sets. convenient to use the continuous updating type weight matrix Wn{6), consistent estimate of asymptotic variance of n~''^ ^"_j mi (9). 10 i.e. to let Wn{0) equal a Since Qn(^) approaches Q{0), borhood of 0/. The we should be able to that 9 tell ^ Qj \i 6 is outside some neigh- neighborhood depends on the rate at which the boundary of 9/ can size of this be learned. In the remainder of the paper, the rate we consider the parametric rate and hence the is points of uncertainty are those 6j that are within a 1/ \/n neighborhood of the true set 0/. Such points are of the form + Oi In order to keep track of such points central interest is The over a suitable domain. dQj. Given 6j eQi,Xe for Oi M'^. will suffice to record all pairs of the it form {9j,\). Hence of the local empirical process (^/,A) set XI Vn, ^ in{9i,\) ~ n(Qn(.6i + \l^) domain pertinent Vn(ei) := {A e R'' : for A + A/Vn' 6>/ Q{0j)), be shown to be the boundary of identified for dj will £ dQi, the pertinent domain - is € 0, for all n' € [n,oo)}. In addition, the following limit version of Vn{dj) will play an important role: V^{9l) := {A G R'^ : 6'/ + A/Vn G for all sufficiently large n}, (3.8) where when 9i^ int(0), V^{9i) = R'^. Thus, V^{9j) plays the role of the limit local parameter space relative to interior of the parameter space, parameter space, i.e. 9j i.e. 9i £ int(0), V^{9[) R'^. When 9i is When 9j is in the on the boundary of the S 50, the local deviations A should be constrained to the local parameter space V^{9j) of the specified form. This situation case, as characterized in = 9j. Andrews in the partially identified models is similar to the one arising in the point-identified (1999). Unlike in the point-identified case, the is more an exception. of a rule rather than boundary problem It arises even in the simplest leading cases, as will be seen in Section 3.3. Another important set is the subset of the local parameter space Vn{di) where the local deviations A are constrained to be towards the interior of the identified A„(6i/) := {0} UiXeR"^ -.91 + X/Vn' set: € int(0;) for all n' In addition, the following limit version of An{9i) will play an important Aoc(^/) := {0} U{XeR'' Note that since int(0/) C 0, is A„(6'/) -.91 is + A/i/n e int(0/) 11 [n, oo)}. role: for all sufficiently large a subset of Vn(0/), which also a subset of V^{9j). G is n}. a subset of V^{9j), and A^{9i) Given above = Cn ^ definitions, i shall obtain the following representations of the coverage statistic Qn) supe^ge, 4(6'/, 0) en{9i,0)-Mgj^^/^^Q \ snpg^^Q^ ^ - sup n (QniOl) f we s^Pe!ede,,\eAn(ei) ^n{Oi, A) \ ^^'Pe,eaeI^e^r^{e,) ^n(^/. A) In this representation, the first + - en{0i,\) if g„:=0, if Qn .= mUeQ Qn{e), Op(l) if 0, ifng„74pO. + Op(l) infe,ee,,Aev„{9,) 4(6*/, A) ng„ -*p equality follows by definition of the local empirical process £„(^, A), To guarantee non- while the second equality emerges from the analysis given in the appendix. degenerate asymptotics for this statistic, we will require that £n{Oi,X) converges to a sufficiently well-behaved random element looi^i, A), in the finite-dimensional sense coupled with some uniformity, which we will call the This form of convergence quasi-uniform convergence. will be sufficient to establish that ^ _^ ^ .^ \ s^Pe,ede,,xeAocie,) ^oo{di,X) The {9j 1 — - inig,^Q,^xev^{e,) ^oo(6'7,A) if ng„ y^p 0. following conditions ensure that this convergence takes place. Assumption C.l 0/ ifng„^pO, s'^Pe,ede,,\eAooie!) ^oo{0i,\), \ ^ dQi, for any G int(0/) e. : When (Degenerate Asymptotics on the Interior). e inigi > there exists large ^qq^ Since in{Si, A) > \\dj — 0, this 0'j\\ > enough S > and int(07) not empty, so that enough n such that for large = 5/^/n}, supg^g/^^^j Kn(^/,0)| is /„((5) :— with probability no less than implies that nqn = in(Or,X) inf = wp ^ 1. eiee,MVn{ei) Condition C.l is motivated by the fact that in interval regression examples £n{9i,X) We = wp -> 1 for any fixed di for large n G int(0/) and A G W^. provide examples and further discussion of this interesting phenomenon in Section 3.3. Condition C.2 puts further convergence conditions by appropriately matching the large sample behavior of function in{0,\) with that of some function ^oo(^, A), which we limit of in{9i,X). In order to state the condition, define the following u„(A) := sup ln{di, A) and e,edei /ri(A) := inf in{9j, A) for is the quasi-uniform key functionals n < oo and n = oo. e,edej Assumption C.2 (Quasi-Uniform Convergence Near Boundary). K call any compact subset ofW^. Then 12 Let 9j G 50/ and A G K, where i. £n(^/)A) > which continuous in X for each (a) ii. is is ifQi converges weakly in finite-dimensional sense to some function ioo{Si,X) = stochastically equi- continuous; ifQj to 0, dj. 50/; u„(A) converges weakly 7^ > Uao{X) in finite- dimensional sense, and (b) Un{\) to dQj, then (a) (uca{X),looW) in finite-dimensional sense, and jointly converge weakly {un{X),lnW) (un{X) and ln{X) are stochastically (h) equi- continuous. supg^^sQ^^xeAo^ie,) ^oc (Oi, iii. < >^) oo a.s. Condition C.2 extends the standard conditions used to derive the asymptotics of criterion func- where 0/ tions in the point-identified case, (1994). Although asymptotics near the boundary is be non-degenerate. to and for verifying A.S. C.2-(i) requires that converges in finite-dimensional sense to some limit ^co(^, inf transformations of £„(^/,A) over imposes the conditions required C.2-(i.-ii.) for obtaining the limit distribution of coverage statistics •^ra(^i-^) Newey and McFadden e.g. degenerate, as condition C.l states, the asymptotics in the interior is assumed a single point Oj, see is 90/, denoted Un{X) and C.2-(ii) requires that '^)- sup and ln{X), converge in finite-dimensional sense to respective sup and inf transformations of ^oo(^/>-^), denoted as Uoo{X) and looiX). and requires that u„(A) hmit coverage Note that Z„(A) are stochastically equicontinuous. statistic C, since snpg^ ^qq^ xeAa^idi) ^oo (^7) A) will in C.2-(ii), it is K is is be one of the components of stochastically equicontinuous over fails interval-censored outcome), while it Assumption C.3 least as large as uniformly in {Oj,X) such that iy{6j,X) positive constants Ci is and S of interest (the regression holds in others (the generalized (n{di, A) Condition C.3 some cases to hold in (Local Quadratic Bound). For any such that with probability at I — C. • • of model with moments). depend on i>(6i,X) 5f = inf^'ge^ ||#/ -H Xj \fn ~ &|||, for set estimates, and along with C.l and C.3 extends, to the standard conditions needed to obtain the rate of convergence of extremum Theorem 13 for e. C.2 allows us to verify the main previous high-level conditions A. 2 and A.S. timators in point-identified case; see conditions in K 0, there is a sufficiently large positive min[;y(6'/, A), > K/^/n, where that do not > method C.2-(ii). for large enough n e > Ci n e needed to obtain rate of convergence set-identified case, the the 90/ x K, any compact subsets of K'^. This stronger and simpler condition certainly implies However, this stronger condition some C.2-(iii) insures tightness of possible to require that (6 J, A) H^ in{9i, A) where It also 5.52 in Van der Vaart (1998) and es- Newey and McFadden (1994), Theorem p. 2185. 3.1 (Limits of C„). Suppose A.l and C.1-C.3 hold. Then A. 2 holds, in particular, \ supe,eae/,AsAoo(e/) ^oo{Oi,X) -in{e,ee,,\ev„{0,) ioo{Oi,X), —>p where nqn Theorem may be is necessarily occurs if qn '= 3.1 verifies easily checked in Assumption A. 2. The theorem dQi and inf in Q. under a statistic attains set of conditions, which a limit distribution, which and sup transformations of the quasi-uniform limit foo(^/,A) we would like to know certain properties of the level sets. 3.2 (Rate of Convergence and Sandwich Property). Suppose A.l and C.l C Then, for some positive constants I. non-empty the local parameter spaces A^{9j) and Voo(^/)- In addition to this basic result, Theorem is states that examples of interest, the coverage determined by the appropriate over or i/int(0/) ifnqn/*pO For any k —p oo such that k - C.3 hold. and C' = Op{n), Qj C Cn[k) C G^" wp -^ 1, where En = C yk/n, so that < C \Jk/n wp dH{Cn{k),Qj) II. For any k G [co,ci] In?!, the Qj C and, provided & — > C„(fc) C 9;" wp -^ ^ at sup Qbi9) 3.2 verifies Assumption A. 3 vergence of the level sets to the identified metric, is 1, where Cn polynomial - sup = C hin/y/n, rate, Qb{9)) = Op(l). 6ee, eeS;" Theorem 1; sandwich condition A. 3 takes place, oo and b/n 6( -> for parametric models and characterizes the rate of con- set. The essentially Xj \fn. 14 rate of convergence, as measured by Hausdorff Analysis of Regression with Interval-Censored Outcomes. 3.3. Xi = 1 and suppose E\Yi] < £[¥2]. Then Qj = {9 <0< E[Yi\ : First consider the case £[¥2]}, that is 9/ = when [£;[yi],£;[y2]] Then Qn{9) = s2 [Y^-eY_^ „n2 /,-, , + {Y2-ey_ and Suppose that V^iXi - EYi,Y2 - EY2)' ->d {Wi, W2)' ~ A^(0, fi). Then Udi, A) = and the finite-dimensional ^ wp limit of {ln{EYr,y), in{EY2,\))' is Therefore the finite-dimensional limit of £„(^,A) ^oo(^/, A) Theorem in{di,X) = [Wi - 3.3 verifies C.1-C.3 for this . G {EYuEY2), 1 if 0/ \)li{ei = is ((W^i-A)2_,(Ty2-A)2_)'. given by EYi) + {W2 - xtnei = example so that -^00(^7, A) is EY2). also the quasi-uniform limit of This implies by Theorem 3.1 that Cn^d C= sup = max £oo{Oi,X) [{Wi)l {Wit] , . e/€9e,,A6Aoc(«'/) One can use the above distribution to obtain the critical values and corresponding level set with the required coverage property. Before proceeding to a more general case, notice that the inferential strategy developed in this simple example can be used in other problems where 0/ F\ <9 < is is interest. All G such that is a set of one needs to do in this class to derive the joint asymptotic distribution of the sample analogs of the endpoints of the interval of interest and obtain via simulation the critical values for the variables, as outlined above for the case when Fj = EYi and F2 = EY2. Getting back to more general settings, suppose that (3.9) all F2 where Fi and F2 are functionals of the distribution of the observed data. This models that provide interval hounds on the parameters of of models defined as the set of X £ {xi,...,xj} 15 P-a.s., maximum of two random where the 0/ is first component determined by the It is X Condition (3.9) assumes that is 1. eM."^ > x'^e : assumed that 0/ C 0, where X n and [xj) x'^d < t^ {xj) a compact subset of is (3.11) < := {xj,j for all j 90/ = and int(0/) is empty Define ti{x) = The identified set {Oi e 0/ 0/ in R'^, so that E{Yi\x) and T2{x) " (3.13) = n{xj) =n x'^ej : = if J}. J) has rank d, {xj) or x'jOj = 50/ < and that the d x J matrix IR"^, which rules out redundant parameterization. Then the boundary of the (3.12) discrete. is set of inequalities: 0/ := {e (3.10) X of = and only ti (xj) if some for , = ti{xj) identified set j < T2{xj) for dQj is J}, some j. E{Y2\x) and consider the objective function '=1 E ^^' = fc i>2, j = l,...,J. i:Xj=Xj In this case for := ^ "121=1 ^-^i rij — ^j]> J ^riJ2^ " ln{9i,X) {{n{xj) - - x'j\/V^)l + x'^9j ihixj) - x)e, - x'^X/V^t) i=i Define Wij := y/n{Ti[xj) - ti{xj)) and W2J := \/n{T2{xj) -T2{xj)) for j = 1, ..., J. Assume also that a central limit theorem and a law of large numbers apply so that {mi,W2i),...,iWij,W2j))' -^d {{Wn,W2i),...,{Wij,W2j))' —»p Tij/n Theorem for pj each j = 1, ^ J^{0,n), and J. ..., 3.3 (Interval Regression). Let the basic conditions specified in Section 3.2 hold and also Then C.1-C.3 and A.1-A.3 (3.9)- (3.14) hold. are satisfied. Moreover, J ioo{9i, A) ^ ^ f = J2pj [(^U - x'^X)ll{x'j9j = T,{xj)) + {W2j - x'^Xf_l{x'^9j = wp This theorem orems — > 1 occurs verifies 2.1, 2.2, 3.1, and when int(0/) T2ix^))] i/ng„ -+p supe,gaQ,;^gA<^(g^)^oo(6'/, A) \ s^Pe,ede,MAc^{9,)^ooi9i,X) -migj^Q,^XeV^{e,) ^oo(^/,A) where nqn = is non-empty or if qn ifnqny^pO '.= 0. Assumptions C.1-C.3 and A.1-A.3, which implies that the 3.2 results of The- apply to the interval regression case. Therefore the inference procedure 16 proposed Theorems properties given in The Further, the confidence regions have the stochastic in Section 2 is valid in this case. and 3.1 3.2. distribution of the limit variable but this does not create a problem simplified much C is not pivotal, and has no for the inferential further, unlike in the trivial case with known method proposed Xi = One can I. closed analytical form, C can not in Section 2. simplify the leading term as J sup = ^oc(^/, A) Vp, \{Wi,)ll{x]ei = TiiXj)) + sup Analysis of the Structural ple, and then generalize it. = T2{xj))] , -" ^-^^ but other terms do not appear to simplify 3.4. {W2j)U{x'.ei l S/eae, e,€ae,,A6Aoo(fl/) above expressions. in the Moment Equations Model. We first work out a simple exam- Consider the usual two-stage-least-squares model which leads to the following objective function Assume < that r = rank EXZ' < dim(^), so that the identified set linear subspace of R'^ intersected with 0. Q, Note that 0/ has empty when there is weak = {e = eo 0/ + S is Qn{9) is of an r-dimensional defined as follows: for any ^o such that Ee{9o)Z ee such that 5' EXZ' = pivotal and can be inverted Suppose is for simplicity that g„ is AniOl) = (a„(0/) = -^ + no longer pivotal. := (which is also true when q-a := inf9ge(5n(&) but dim(Z) A'- Y, ^^^0 (- E ^'^0 i=l i=l Y.{Yi - XleI)X^ = 4= E ' [^r.{Bi) + A'- Y. XiZl) , i=l ^ii^I)Z^ Under standard sampling conditions n - Y 1=1 ^ ZiZ- -^p EZZ', - In the empirical dim(A')), so that ir,{eu A) 0, in the case for confidence intervals. the partially identified case, the statistic most relevant to making inference on 0/ process {Qn{Si),di £ ©/) which = 0}. Under the usual circumstances, even interior relative to R'^. identification, Qr consists n Y 2 ^i^'i -*P EZX', and =1 17 {e{ej)Z, Oi e 0/} is Donsker. < Hence the finite-dimensional and quasi-uniform ^00(^7, A) where {A{9i),9i e 6/) under is iid a zero We is = weak given by is + X'EXzJ (^EZZ'Y^ (^A{9:) + X'EXZ'^' (^A{ei) the limit of £„(0/,A) limit of the empirical process {An{Oj),d] € 0/). For instance, sampling, provided that {e{6i)Z,9i € 6/} has a square-integrable envelope, the Umit A(-) mean Gaussian process with covariance kernel given by EA{9j)A{9'j)' " T 1 = An{9iy(^"rt i=l Notice that the limit is Ee{9j)e{9'j)ZZ'. and thus therefore conclude that C.1-C.3 are satisfied C„ = -1 An{9j) -^d C ZrZ'^ = sup A{9 1)' [E Z Z]'' A{9 j) e,ee, ^ not pivotal and depends on knowing Q/}^ Also, compactness of necessary condition for the limit variable C to be 0/ is the finite. Next, we generalize our method to the general nonlinear method of moments. Let the following partial identification condition hold: This is > C \\Emi{9)f (^l^-. C There are positive constants min[ ini and such that S uniformly for 9 € 0. \\9-9'j\\,Sf, both a partial identification condition and a smoothness assumption. This condition implies that Emi{9) = if an only ii 9 £ Qj, and also imposes smoothness on the behavior of Emi{9) for points 9 near 0/. In the point-identified case, global identification and the VeEmi{9) near 9] ordinarily set identified case, the imply Jacobian (3.15); see e.g. may be Theorem is : positive eigenvalue of Theorem hold and EZ'Xv = In Hence if rank pj — ^|| > EZ'X is iEX'Z){EZ'X), we have \\EZ'X{9 - Pakes and Pollard (1998). In the 0, the vector {9 non-zero, for 9})\\'^ > Cq = EZ'X{9 — 9]), — 9}) is orthogonal Cq denoting the minimal \\9 - 9*j\\'^. 3.4 (Generalized Method-of Moments). Let the basic conditions specified in Section 3.1 let i. 0/ = 90/, continuous Jacobian G{9) (A(e/)' 0}. 3.3 in IV model we have that Emi{9) the closest point to 9 in 0/. Provided that to the hyperplane {v rank and continuity of the Jacobian degenerate, which necessitates a statement of a more careful condition (3.15). For example, in the previous linear where 9] full the case + ySXZ') ii. = 6e o S/eEmi{9), Wn{9) > dim(z) [EZZ']-^ (A{e,) Cn Donsker {mi{9),9 € 0} v. dim(2;), and class, Hi. = W{6) + qn = Emi{9) Op{l) uniformly in 9, where infgge Qn(S), + X'EXZ') and ^d C= , sup e^{ei,0)B,eB, 18 satisfy (3.15) inf ioo{9i,\)o,eei.>.€Vocie:) then and have W{9) im{9i,\) is positive definite and continuous for t^iOj, A) = (A(^/) Then, C.1-C.3 and A.1-A.3 hold, and all 9. + X'G{e,))' W{ej) {A{9i) + X'G{9i)) A{9j) i—> weak the is —>p nq-n necessarily occurs if qn := The theorem orems Section 2 in is and 3.2 = 3.1 and 3.2. 4. We We illustrate -£'[n~"' = inf^ge Qn{9) X^"_j ?nj(^/) ^"_^ but dim{mi{9)) < mi(^/)']. Above, dim(^). GMM case. Therefore the inference procedure proposed in Further, the confidence regions have the stochastic properties given Similarly to the interval regression case, distribution of the limit variable does not create a problem qn if lim„_oo C.1-C.3 and A.1-A.3, which imphes that the results of The- apply to the valid in this case. Theorems or verifies the conditions 2.1, 2.2, 3.1, a zero-mean Gaussian process limit of the process 9j i—> An(0/), with the covariance function £'[A(^/)A(^/)'] ifnqnT^pO ^oo{^/,A) suPe;ee,,AGAo„(fl,)^oo(i9/,A) -infe,ee;,AeK»(e/) where 9j , C not pivotal, and has no is known method proposed for the inferential C depends on 0/. Hence the closed analytical form, but this in Section 2. Computation and Empirical Monte Carlo our methods above with an empirical Monte Carlo that uses data from the CPS. also provide details on the computational method that can be used to construct the confidence sets. 4.1. Computation. An subsampUng method issue in the is being able to use an algorithm to construct level sets of an objective function. Ideally, one would of grid points and evaluate the function on those efficient like to use a rich set points. However, as the dimension of 9 increases, constructing a simple uniform grid becomes computationally infeasible. The Metropolis-Hastings algorithm provides a computationally attractive method for generating adaptive grid details of the numerical approach we use = is 1. Generate a grid of points Q 2. Given a starting critical value objective function as: 3. C„(k) = At each subsampling stage j : j Cj_b,n = 4. Compute the Q-quEintile c(q) of 5. Then compute C„(e(a)) ''This algorithm is {^g and Cq G B : = k=Colnn, we can compute the level set — ming ec„(i[) Qn(^g)) < k}. l,...,Bn, (Cj^b.m j = - 1, in '^ of the compute Qb(^g) for all ^gGCn(k), and the Qb(6'g) - mine^gc.{k) Qb(Sg)) • ••iBn} mine^gc„(k) Qbl^^g)) < c(q')}. a valuable method of generating grid points adaptively, so that they are placed in relevant regions only. See Robert and Casella (1998) for a detailed description; and Chernozhukov and examples The in the following algorithm: b(max9^gc.(k) n(Q„(6'g) sets. using the Metropolis-Hastings algorithm. n(Qn(&g) where subsample coverage statistic, e.g. = summarized {9i, ...,9^^) {^g G numerical non-likelihood settings. 19 Hong (2003) for related An important practical issue consistency of estimate Cq is is the choice of the initial point not affected by the choice of cq as long as cq Hence a reasonable procedure (2.2). For example, in GMM for selecting cq how 2.2 suggests the bounded is in the sense of can be based on pointwise testing procedures. a pointwise critical value for testing Q{9) of a chi-squared variable with degrees of freedom equal to the section for Theorem cq. to choose cq in the interval regression case. of the asymptotic distribution of the coverage statistic More = at ^ number is given by the a-quantile of equations. See the next generally, one can use the a-quantile computed under the assumption that 9j is a singleton. We examine the actual finite- Empirical Monte Carlo: Returns to Schooling in the CPS. 4.2. sample performance of the inferential procedures Population Survey. Our "population" wave of the proposed in this paper using data from Current a sample of white is CPS. The wages and salaries men series are not top ages 20 to 50 from the March 2000 coded or censored and so we are able to use this population to construct 1. confidence regions for the returns to schooling parameters in the point identified case, and 2. confidence regions covering the identified set Our Monte Carlo summary when we artificially based on random sampling from the original data is statistics for bracket the data. set. Table 1 below provides our data (population). The "true" returns to schooling coefficient Table 1. Variable Wages and Salaries Education Population Summary is .0533 Statistics Obs Mean Std Dev Min Max 13290 66667.6 51968.41 1 513472 13290 11.77 1.89 1 16 with a constant term of 3.91 obtained from a least squares regression of log of wages on education in the "population" 4.3. 0/ . We by describing the start first details of the computational procedure. Starting values and Implementation Details: The algorithm used to obtain estimates of is the same as the one described on page 20 above. 1. We draw a sample of size The steps of the procedure are as follows: n from the above population (this population will be artificially bracketed below). 2. We build an initial estimate C„(co) at the starting cutoff level cq (choosing cq is described next). 3. In the subsampling step, we obtain the consistent estimate, Ca of the cutoff level by subsam- pling the coverage statistic. 20 At this point One may In The all we levels that we For example, in the method of moments case, the initial level set of critical value the objective is = the following way. Let Yf^ -^{Yu + function Qn{d) applied to the data {Yii,Y2i,Xi,i point-identified at some value 6^, and for < Yu = and Y2i) is problem specific. from a x^ with an appropriate degrees of freedom can be used as the starting value. In the interval regression example, value Co get are very close use the numerical procedures as described in Section 4.1. used to construct the is step 3.^^ Calnn and then repeat iterations. of the above steps starting value cq that = cq but we find that the cutoff iterate several times, no or at most two after 4. one can go back to step 2 above and set Y2i = Vj", we choose an initial then we use the objective This generates an auxihary model, which n). which we can compute the is limit distribution of n(Q„(0g) - Q„(eS)), and take which its quantiles as the starting value a consistent is method cq. In what follows, we used subsampling to estimate of estimating cq by Theorem 2.6.1 in Politis, Romano, and Wolf cq, (1999). Properties of the Set Inference Procedure in the Point-identified model. Here, the 4.4. data on v/ages is not interval measured (the model point identified), but is we nonetheless apply our procedure to obtain the confidence region. The objective function that we use is minimum the distance objective function J=16 Qn{e) (4.1) = Y. i=l ^(fi(x,) " x'^e)l + (f2(:r,) Since wages are not interval-measured, we have that f\{xj) obtained are compared to the the joint confidence approach. In Figure we provide graphs 1, subsampling procedure In Table sample 2, sizes: n = 1000 and n b = 4.5. for the case sizes. where n x'^e)l. T2{xj),j = I,..., J. obtained using the usual n = 400 and n = 2000 The Wald results inference observations and see that the close to the eUipse obtained using the usual chi-squared approximation. we examine the coverage sequence of subsample and is for ellipse = - = = 2000, using As we can 1000, it We provide the coverage for two simulations. We report the coverage for a of our inferential procedure see, R = 600 . coverage seems monotonic initially in the subsample size peaks at 6 = 200 while for the case where n = 2000, it peaks at 500 and comes close to 95%. Properties of the Set Inference Procedure in the Set-Identified Model. To examine the identified set of parameters in the case of censoring of the dependent variable, we bracket the income data into 15 different categories. These brackets are In the interval regression example and similar situations, In 21 (in thousands) n may be dropped. Figure Point Identified Case: Subsampling vs x 1. C g Subsampling 0,065 Ellipses Ellipse 0.06 S - Chi Squared Ellipse UJ , 0.055 3.4 3.6 Constant Table 2. Finite-Sample Coverage Property in the Point Identified Model Subsample b=50 b=80 b=120 b=200 b=300 85.1 88 87.2 93.1 91.2 b=200 b=300 b=400 b=500 b=600 86.1 88.2 90.5 95.01 93.3 n=1000 Coverage (^95%; n=2000 Coverage [0,5] [5,7.5] [30,35] , , [7.5,10] [35,40] , , [40,50] (<dh7o) [10,12,5] [50,60] , Notice here that the topcoding obtain 718 observations in the , , [12.5,15] [60,75] is artificially first random without replacement from artificially [15,20] set at , [20,25] , [25,30] [100,150] , [150,100000] , $100 million. To give a flavor of the data, we bracket, 1211 belong to [50,60], 1598 belong to [100, 150] while through our 95% confidence region C„(c. 95) Using the , [75,100] , 734 are making above 150. The objective function at Size for is the same as the one used sample sizes = n earlier. In Figure 2 600,2000,4000, and 10000 drawn the original "population". bracketed population, we can obtain the identified set by collecting the parameters that satisfy the following set of J inequalities corresponding to the set of the level of education takes: E[y\\xj\ < 60 + biXj < E[y2\xj], 22 j = 1, ..., J. J values that We then graph the identified set 0/ along with the as the sample n= 95% size increases, the set estimates shrink 600, our set estimate of the intercept Figure C„(c.95): 2. 9/ is confidence region Cn(c. 95). towards the identified [3.2,4.71] while the true range is Figure vs n=600, b=60. Cn(c.95). Notice that For example, at set. [3.6,4.4]. 9/ 3. vs n=2000, b=200. 0,18 0.16 0.14 Ik. V ^^k- — Subsampljng ConfiOence Set c 0-12 D UJ (A E E 5 0.06 |oos Subsamplinq Confidence Set ^^. So.OB -^ 0.0a — ^^ "^^^j^— 0.04 0.02 Sel Constant Figure Figure Cn(c.95). 4. 9/ vs 9/ 5. Cn(c.95). n=4000, b=400. vs n=10000, b=2000. 3 0.06 a: Conridencfl Set 3 3.5 4 4.5 Esbmale 5 Constant To calculate coverage probabilities, we draw R= 600 random samples from the population and record whether our set estimates using these samples cover the identified set. For every sample, we construct the set estimate, and calculate coverage in the following manner. Numerically, 23 we store the 5.5 6 Table 3. Finite-Sample Coverage Property in the Set Identified Model Subsample Size N=1000 Coverage ('95%; N=2000 Coverage (<d^%) identified set Cn{Ca)- 0/ and Cni&a) As Table 3 as arrays, and Figures b=100 b=150 b=200 b=250 b=300 84.1 90.2 91.3 88.5 82.8 b=200 b=300 b=400 b=500 b=600 86.34 90.1 92.2 94.3 91.2 and tlien clieck wlietlier all the points in 2-5 show, the coverage set estimates converge to the identified set as the of our inferential methods supports the samples This paper provided confidence regions are contained in similar to the point identified case and the size increases. The numerical performance large sample theory developed in 5. The proposed is 0/ Theorems 3.1-3.3. Conclusion for identified sets in inference procedures are criterion function based, models with partial identification. and our confidence regions are certain level sets of the criterion function in finite samples. In the case when the model is point-identified, our confidence sets reduce to the conventional confidence regions based on inverting the likelihood or other criterion functions. The proposed procedure was shown to be valid under general yet simple conditions. Along with inferential procedures, we have developed methods of analyzing the asymptotic behavior of econometric criterion functions under set identification the rates of convergence of the confidence regions to the identified set. regressions with interval data and set identified method of We and also characterized applied our methods to moments problems. We also assessed the performance of the methods in Monte Carlo experiments based on the Current Population Survey data. We found that the methods perform well and in accordance with the asymptotic theory developed. 24 Appendix We A. Appendix use the following notation for empirical processes in the sequel: for 1=1 E E denotes expectation and wp — , (-B/(l^.))/=/- P* and P, and corresponding notions probabilities, under P* X) denotes expectation evaluated at estimated functions /: van der Vaart and Wellner (1996). in distribution {Y, 1=1 * Ef{W,) = Outer and inner W= » —>p means "with the inner 1 of weak converegence are defined —j denotes convergence in outer probability, and "with the inner probability no smaller than 1 — e for probability approaching 1," sufficiently large n" A . as in means convergence and "wp > table of notation is 1 — e" means given at the end of the Appendix. Appendix Lemma Proofs of Lemma B. B.l. Proof of B.2. Proof of Theorem 2.1. Proof B.3. Proof of Theorem 2.2. Step 2.1. Proof (B.l) 2.1, Theorem 2.1, and Theorem 2.2. D given in the main text. is is I. D given in the main text. We Cj,6,n have that = sup ai,{Qj,b{d) - Qj.b) , esc„(c) where qji, := if Qn = and Qj^i, := infgge Qj,b{S) otherwise; Qj,b{&) denotes the criterion function defined using the j-th subset of the data only. Define B„ - B-' J2 nCi,6,n Gb,n{x) (B.2) s„ <X} = B'' Y, <X' {Cj,b,n " Cj,b.n)}, HCj,b,n j=l i=l where (B.3) Cj,b,n = sup at {Qj,b{0) - qj,b) ese, In what follows, the main step is to show that Cj^b,n can be replaced by This Cj^b,n- will be possible due to the sandwich assumption, and despite that the constant k used in construction of the preliminary set estimate Cn{k) is By data-dependent. A. 3 wp — > 1, for some deterministic 0^" we have e,cCnCk)Qe]\ (B.4) where 6^" := {0i +t: (B.5) set \\t\\ < en,9i G Qi}- Hence sup ab{Qj,b{d) - qj,b) see, < sup wp - 1 for all ab{Qj,b{d) - < B„ i qj.b) eec„(c) V '^''''' ' > < sup ab{Qj,b{0) - gj,b)- see'," ^ C,,t,„ 25 ' > ^ Cj,i,,„ ' Hence wp — > 1 B„ a, Js) = B-' (B.6) By A. 2 C6 ^d ^ B„ <x}< l{C,-,b,„ < G6,„(i) Gi,,„(a:) s B"' ^ l{C„b,n < x}. C and by A. = C6 (B.7) + 0p(l)^d Ct C, so that by Step II Gi,„(a:)-^p (B.8) if X is G(2;) := P{C < G6,„(j;) ->p i}, G(a;) = P{C < k}, a continuity point of G{x), which proves that: G6,nW--p G{x) (B.9) at the continuity points x of G(x). Furthermore, (B.IO) if G is continuous at c^ = G-l{a)^j, c^ = c^ = G'''(a), G-'{a), since convergence of distribution function at continuity points implies the convergence of quantile functions at continuity points. Finally the claim I, that P(e/eG„(c,))-»a, (B.ll) follows by Theorem Step II. 2.1. This part shows (B.8). Write Br, _ = P{Cb<x} + o^{\) = P{C <x} + at the continuity points x of G{x) = P{C < (B.13) where i}, as long as > n (a) follows and the 6 — » n— CO, Y^rl^Y. 1(^^.^- by the variance bound (B.15) 0, > oo; from (B.14) in Politis, Op{l) for bounded Romano, and Wolf < (b) follows Cb^d and a-mixing by C, as definition of convergence in distribution on M. 26 o (1) j [/-statistics for i.i.d. series (1999); and = :r} b^ oo, series given on pages 45 and 72 Likewise, conclude B„ _ = B-'Y,l{Cj,b,n < Gb,nix) X} '"' (B.16) = x of G{x) at the continuity points Step Wp ^ III. = P{C < = P{C <x} + Op{l) x}. e,cCn(k)^h{e,,CnCk))=0. wp -^ 1 =^ Cn{k)ce'- (B.18) Hence Op{l) 1, (B.17) Then, P{Cb < x} + wp — h{Cnik),e,)<hie''',e,)<en. 1 > d//(e;,C„(fc))<e„. (B.19) Thus Claim II is D proven. Appendix Proof of Theorem C. 3.1 First recall that Cn := sup n{Qn{di) - q-n) e,se, ^^^^ ^ We f supe,ge, 4(5/, 0) if g„:=0, \ supe^ge^ £n(6i7,0)-infe^+;^/y;;ge 4(6li,A) if gn := infege seek to establish that Cn -^d C, where /Q^N ^ _ ifng„->pO supe,g3e^;^gA„(e,) ^oo(6'/,A), I \ supe,gge^,;^gA«(9,). ^oo(6'/,A) -infs,ge,,Agv^(9,) ^oc(6'/,A) AooC^*/) := {0} (C.3) U {A e Ko((9;) := {A (C.4) Step I: Case when ng„ -^p 0. G R"^ K"^ 61; : : Since nqn —*p + A/^n + 5/ in what follows Then we would we e int(e/) A/v/n G 9 if ng„ for all sufficiently large >p 0, n}, for all sufficiently large n}. we have that 0, c„= (C.5) Hence Qn (5)- sup e„{e,,o) + op{i). ignore the Op(l) term. like to show the following simple, but important representation: C„= (C.6) sup 4(e;,0)= sup 9iee, 4(^7, A) e,£ee,,>ieA„(«;) where An{ei) := {0} (C.7) To prove 6'i + (C.6), X* /y/n = we need 9i. If 9/ to U {A e show that R'' for : 61/ + A/Vn' £ int(e/) for all n' € [n, oo)}. each 5 G ©/, there exists 9} G 50/ and A* G An(&J) such that G 96/, then simply take 6) = dj 27 and A* = 0. If 9: G int(0/) take a sufficiently small ball centered at 6i of radius 5 > , denoted Bs{Oi), such that Bs{9i) C mt(0/). Then start expanding the radius of the ball 5 so that to find the minimal value of S such that the boundary of the ball includes a point G d&j. Such radius 9} Oj all belong to int(6/) construction In what exists by compactness of 0/. Then the points on the G (96/ and A* G An(^J). Hence (C.6) 9*j follows, we is \* /y/n/, n' £ = 1. G/ has an empty 2. 0/ has a nonempty interior relative to 0, so that 0/ By Cn= interior relative to ©, so that sup 4(6'/, 0)= is sup (9/ — O} 9*j)y/n. and By j^ 90/. sup ^00(^7,0). B,eae, A^{9,) = {0}, also rewrite (C.8) using general notation Cn= sup sup en{9i,0)= Biee, 2. = 90/, C= en{0i,O)-^d = An{ei) Case segment between where A' empty, (C.9) (CIO) 0/ e,ede, 9,eei we can line C.2-(i) {C.8) Since int(0/) oo), distinguish two cases: Case 1. [n, true. Case Case so that + and can be parameterized as 9} For any 5 C (^ [9, X) -^d , = sup 9;ese,,AgA„(9,) > ^co(9/,A). e,eaei,xeA,^{e,) decompose Cn=meix[C:{5),Cn{5)], (C.ll) where sup C;((5):= fCl2l ' ^ and In(S) = {9i G int(0/) By Assumption : infg' C.l for any gge; e > 11^/ Cn((5) := - ^/ll > i5/\/n}, as defined in = liminfP.(c;(<5) also that Cn{5) > sup 4(^7,0), 8;ee,\/„(i) Assumption C.l. there exists 5 sufficiently large such that (C.13) Observe and in{Oi,0) o,eui6) by construction. Hence for 0J any e > 1-e. > there exists 5 sufficiently large such that liminfpJc„=C„(5)| > 1-e. (C.14) Observe that Cn{5)= (C.15) sup 4(^7, A). 9yese,,AeA„(e,)n{||A||<5} By Assumption C.2 (i.-ii.) any compact sup (C16) ^ ' set i—» K e.n{e,,-)^ e,eae, ' where A for sup i^{9,,-)mL°°{K), e,ede, supg^ggg^ ^oo(^/, A) has uniformly continuous paths. The next claim Continuous Mapping Theorem that C„(<5) fC17) ^ ' The proof (CIS) ^d C((5) := sup eoo{9,,\). e,eae,,xe\^{e,)n{\\x\\<s} ' of this claim is given below. Hence for any closed set limsupP'{C„(<5) G n—OO F F} < P{C{5) G F}. 28 is that (C.16) implies by Hence by (C.13) and (C.14) for any e > there exists Vimsup P* {Cn e F} n—*oo (C.19) Therefore, taking e — » and J^ — » oo accordingly, it limsupP*{C„ (C.20) ^ enough such that Sc large <P{C{5,)e F} + 6. follows that gF} <P{CeF}, n—too where C := lim C /p91\ = (5) sup icx>{di,^) a.s. The limit C exists in R by the Monotone Convergence Theorem. By construction C > a.s., and by Assumption C.3(iii) C < oo a.s. Conclude that by the Portmanteau lemma (van der Vaart and Wellner (1996), p. 20) and (C.20) that Cn -^d C. (C.22) Proof of Cn((5) ^d C{5). Observe that K„{Oi) C (C.23) and An{6i) / A„(0/), in the sense that An{9i) uz^, n- (C.24) since by definitions given earlier for „„ all A^{d,) no U^=„ A^^;) (C.25) C A„(6'/) > = A^(ei), for is all n > no, all no > a nested sequence and An(0i) = n-^1 u- = ,„ Ar,{9j) 1 — > A^(5/), i.e. A^{e,), 1 n^ A^i9j) and „„ A„(e;) = A„„(^;). Therefore, 4 sup Cn(5):= (61/, A) etede,,xeA„„{e,)n{]\M\<s} <Cn{6)= (C.26) sup 4(6'/, A) 9,ede,,xeA„{e,)n{\\M\<S} 4(^/,A). sup <5;(<5):= e,eae,,xeA„{e,)n{\\x\\<s} and sup C((5):= eoo{Oi,X) e,eae,.xeA„^(Bi)n{\\x\\<s] (C.27) <C{5):= sup ^oo(^/,A). «;eSe,,A6Ac„(9,)n{||A||<,S} By (C.16) and the Continuous C:{5) (C.28) and for (C.29) Mapping Theorem -d any fixed Hq Cn{5) -»d C(J) Let the underlying probability space be denoted as (C.30) C{5) {0,,J^, sup P). Observe that for every ^oo(^/,A) e,ede,.xsA„ie,)n{\\x\\<s} 29 oi e H, is some necessarily attained at compact and A^{6*i{lo)) n G dQj and A* (a;) e A^{d}{u)) D d^iuj) < {||A|| 5} = C{5) (C.31) Moreover as no — (C.32) » That also compact. is {||A|| < 5}, because the set 90/ is is ^oc(6''(w),A*(a;))a.s. oo, AnoiOi) n / A„{ei) n <S} {||A|i < {|iA|| <5} for each 8, e 99/ hence Koi^iH) n (C.33) so that by continuity of A h^ ioo{6i, A), {||Ai| we have C(5) (C.34) Note that to obtain (C.34) we use that exists a.s. <5}y^ A^(9;(u;)) < 5} a.s. a.s. monotonicity increasing in no due to (C.32), so that lim„o_^oo £(<5) and by (C.27) lim no—oo < C(<5) Moreover, for &/(w) defined above there exists a sequence (a;) a.s. {||A|| that C(5) = C(&) / nolim —oo C_[S) is (C.35) A* n Hence by continuity of A lim (C.36) TlQ Now we —»00 i—> loai^i, A) at Mm C((5)> C{5) a.s. X^^Jyui) in An„(6J(u;))n{||A|| < S) such that A*^(cij) — each 5/ G 99/, i^{e]{uj),\n,{uj)) = io.[e'',{u),\'{uj))=C{5)&.s. Tig—'OO F are ready to close the argument. Let P{C{5) < F} *=^ be any number such that P{C{5) real lim P*{Cn{5) = F} = 0, then < F} (2) < liminfP4C„(5) < F} (3) < lim sup P*{C„ (5) < F} """^ (C.37) (4) < lim sup P*{C„ (5) <F} n—*oo (5) < P{C{S} < F} ^ no —'OO where (1) is for any no > 1 P{C{6)<F}, by (C.28) and the Portmanteau lemma (van der Vaart and Wellner (1996), (C.26), (5) by the such that P{C(5) Portmanteau lemma, and = F} = by (C.34) and the Portmanteau lemma. Thus 0, liminf P.{C„(<5) (C.38) (6) is n— oo < F} = lim sup n— oo P*{Cn (5) < F} = P{C{5) < F}. Thus, by the Portmanteau lemma Cn{S) (C.39) Step II. Case when 9/ Case 1. = 99/ and nq^ 9/ has a nonempty -f*j, 0: p. 20), ^a C{5). Consider two cases: interior relative to 30 0, so that 9/ ^ dQi, (2)-(4) for any by F Case Case 1. 0/ has an empty 2. In this case interior relative to 0, so that ng„ -^p Case 2. I has taken care We 0, of. have by definition of £n(^/, A) and Cn, C„= (C41) We 90/. by Assumption C.l (C.40) which Step = 0/ sup ^„{e/,0)- inf (.n{e,,\). need to show that sup £n(6l/,0), (C.42) tn[e,,\)\^A inf e,+A/v^se Ve,ee, sup ^oc(6l/,A), inf ^oo(6'/,A) e/eee,,AeVoo(e,) \e,eae,A6A„{e,) / Define Mn-= (C.43) inf M:= e,sae;,Aev„(e;) ^ool^^/.A). and e„{Oi,\) inf fl,+A/7Hee Mn Below we show only the marginal convergence the arguments of the proofs of Step I and Step -*d M.. The joint convergence (C.42) follows by combining and applying the Cramer- Wold Device. II Write Mn (C.44) = min[.Mn {K),Mn[K% where Mn{K):= <'.(«/, A), inf 9, + A/yS'ee, d(9i,A)>K/v^ (0.45) M^{K):= fl/+A/vnee, inf u(e/,A)<A-/yTr 4 (9/, A), where v[Qu\)~ (C.46) By Assumption C.3, for any > 0, there is some constants Ci > and 5 > ll^/ + A/V^ - 6>;i|. K large enough so that liminfP.{A?„(/i:) n —oo (C.47) for e inf > Ci min[/C2, nJ^]} that do not depend on > 1 - e, e. First observe that Mn{K)= = (C.48) = ^n(^/,A) inf ^'nf (-niOiA) inf ^n(e/,A), e;€Se,,Aev„(8,)n{||A||<A'} where Vn{ei) := {A e M"^ (C.49) Equality (1) in (C.48) is trivial. First, can be represented trivially satisfies eis 6J + ^/nv{9i,X) : 6>/ + A/\/^ 6 0, for all n' by compactness of 0/ every X* j \/n where A* = yJnv{Qi, A) < K. 31 < X. 5/ e [n, c»)} + Xj -Jn such that v(Qu A) Second, every Qi + Xj \/n with < Kj \/n ||A|| < if Equality (2) in (C.48) is more To sliow subtle. {{6j the intersection of two convex sets (by Assumption A.l {9j + X/^,\\\ < note that {9i it is r]Q,9j G 0/}. Note that + (2) + Bk^^{0)) Bj^^^{0)) n is convex) that contain is A", 5/ G 0/} convex and necessarily contains 9[. Hence n 6 = dj, since for every = dj + \l \/n € {9i + 5^/y^(0)) n 0, points on the linear segment between 9' and 9i are also contained {9i + Bjii^{Q)) n 0. Therefore A = s/n{9' - 9i) e Vn{9]) n {||A|| < K}, and representation (2) follows. 9' By Assumption C.2 (i.-ii.) inf (C.50) 4(^/,-) ^ inf Equipped with (C.48) and (C.50), we can show through the use MniK)^^ MiK)~ (C51) The proof of this claim m ioo{9i,-) L°°{K). of the Continuous inf Mapping Theorem that ^oo(5/,A). stated below. is Hence by (C.47)-(C.51), for any e > there lim^infP.{A^„ (C.52) a sufficiently large K{e) such that is = > 1-e. A^„(/i:(e))} Note that ^= (C 53) M exists a.s. in < by C.2, By R inf ^cc (^/, A) = lim by the Monotone Convergence Theorem. Since M < oo a.s., i.e M (C.51)-(C.52), for any is > e {9i, A) > and tight. and each closed limsupP*{7W„ €F} < set F, lim sup F* "-"°° (C.54) (.^z M (K) {Mn{K {e)) e F} + e "-*°° <P{M{K{€)) eF}+e. Letting e — > and K{e) — » oo accordingly, follows that limsupP'{A^„ n —-oo (C.55) By it the Portmanteau Lemma conclude that Mn--d M. (C.56) Prooi of GF} <P{A^ eF}. MnjK) ^i M{K). (C.57) Observe that Vn,{9i) C Vn{9,) C VU6i), for all n > no so that M„{K)~ tn{9i,X) inf er^de,,\&v„(e,)n{\\\\\<K} <Mr.{K)= rC58) inf <M;:(A'):= iniOiA) inf in{9i,X)- s,s5e,,Aev„„(e,)n{||Aii<K} By (C.50) and the Continuous (C.59) Mapping Theorem Mn{K)^iM{K):= for any fixed no inf e,eae,,Aev„„(e;)n{||A||<A-} 32 ioo{9uX), finite at least for A = in and MniK) (C.60) -^d M{K) - C^ inf e,ede,.>~ev^{8,)n{\\x\\<K} {Oi , A) Moreover, as no "^ oo (C.61) Vn„iei) so that by continuity of A n {||A|| <K}/' i—» iooi^i, ^) ^^ {i|Ai| < A^ for each Oj G 90/, each dj 6 96/ \M M(K) (C.62) The n V^iei) a.s. details of proving (C.62) are nearly identical to those given in (C.31)-(C.36), so are not repeated. Let F be any real number such that P{M{K) = F} = P{M{K) < F} = lim n—'oo > lira then 0, P,{M^{K) < F} <F} sup P'{X„(i<:) n—'OD (3) > <F} liminfP,{7\/(„(A') "^°= {C.63) (4) >^mmf P,{Mn{K)<F} n—*oo (5) > P{M{K) < F} ^ no—>oo where (6) is (1) is P{M{K) < by the Portmanteau lemma and (C.60), (2)-(4) by (C.62) and the Portmanteau lemma. Therefore, for for any no > F}, .. by (C.58), any real F (5) n— oo n— oo Hence by the Portmanteau lemma Mn{K)^d M{K). (C.65) Appendix Steps Step I I: and II prove Claim Case when nq^ —>p I and Step 0. III D. Proof of Theorem proves Claim Recall that Q{9i) = 4(e/, A) := n{Qn(0i (D.l) Since ng„ —p 0, C„= the Proof of Theorem 3.1 II. + \/V^) - sup £„(6I/,0) + Q{ei)). Op(l). we have that sup 4(e/,0) e,ee, 33 3.2 and we have that (D.2) (D.3) by the Portmanteau lemma, and such that P{Ai{K) = F} = limsupP*{X„(A:) < P} = \imMP,{Mn{K) < F} = P{M{K) < F}. (C.64) From 1 = Op.(l). Hence since k —»p oo < C„ (D.4) wp fc -^ => 1 K —^ Condition C.3 implies that for any wp —+ ^ 6/ C Cn{k) wp K/k oo with => 1 -^ h{ei,Cnik)) =0 wp ^ 1. there are positive constants Ci and S such that 1 Ci (D.5) uniformly By in (0/, A) n and since Q(9j) definition of Cn(fc) sup + nQr^{0) < 4(e^, a) min[i.(6i/, A)^, 5^] K/^, > such that u{9[, A) (D.6) • where A) i^(5/, = infe^^e; 11^/ ^/\/n - + 6'j\\. = = Op{l) sup + Op(l) < 4(6'/,A) fc. e/+>./v^ec„(fc) eec„(fc) Hence 9/+A/v^€C„{fc) (D.7) Then the claim is that wp — » C^k) C where 0^ := {dj +1 : ||i|| < Then 4fc by (D.9) and by (a) is wp — > (b) is 1, by (D.5), fc Op(l). inf <^ Ci n — oo and k/n -^ (c) is Combining (D.4) and (D.8) it • + A/v^-^;! l^z 1 » » e2('^/^')"'/v^, G 6/}. Suppose otherwise, then wp — for this pair [Oj, A), (D.IO) 4fc c,6[ v(ei,\)= (D.9) + 1 (D.8) where implies4(6'/,A) <fc min[i/(6i7, A)^ 5^] < for some 6i + X/^Jn e Cn{k), >2(fc/Ci)'/VVn. 4((9/, A) < fc + Op(l), so that Ci n-min[i/(5/, A)^, 5^] > C\-n-rmn[A{k / C\) / n, 5'^] > by (D.7). This yields a contradiction. Thus, (D.8) wp — follows that » Step H: Case when ng„ y^p 0. Observe the relationship between the "centered" and "un-centered" dence regions. Let the un-centered region be denoted as before: (D.12) Cnic)~{9:niQn{9))<c], and the centered version be denoted as (in this proof only): Cn{c):= {e:n{Qni9)-qn)<c], where g„ := (D.13) inf Observe that the following key relationship between the two sets (D.14) (D.15) Cn{k) fc — » oo but k = Op{n). From = Cn{k + nqn) the proof of Mn = nqn = for all Theorem 2.1 Op(l) since /c> 0. we know that Mn -^d Hence (D.16) true. dH{ei,Cn{k)) < h(Cn{k),ei) < /i(e'/'^^'>'''/^,e/) < 2ik/c,y^^/v^. (D.ii) Let is 1 k:=Ck + nqn) = 34 k-il + Opil)). M. Q„ (61). confi- By (D,14) (D.17) CrrCk) Hence by (D.ll) and (D.16) = wp — follows that it ^H (^(fc), 6/) = dnCC^ (fc), SO Cn{k), and the claim now Step 2{k/C,f'^l^< (2(fc/Ci)>/VV^) II, that for k G [cq, ci] Inn, • 0/ C Cn{k) C ej", where (D.19) 6( wp — » — {^! We + * • IKII ^n>0l £ ©/}. and En —> — have that k 6 [co,ci] wp — -Inn > 1. sup Qi{e) - some C> 0. = Then, using Q{6i) < 6( sup Qb{0) - sup Qb{e)) a sequence of positive constants. is wp -> 1, I where Cn =C In sup QhiOi)) sup < + \/Vb - biQbiei Qie,)) - sup HQbie,)) - sup = max[ 4(6/, A) - sup . sup =max[ sup 4(9/, A), sup 4(^/,0)]- sup li,{9[,0) e,ee, e,ee, 4(6'/,0) + = max[ sup in e;se, e,eei sup 4(6'/,0)] 4(6'/, 0), 8;eaej e,ee/ sup 4(^7,0) + Op(l)- sup etee, (a) follows sup ^(e/.O)] - sup 4(^/,0) Op(l), 9/eae, = ib{0i,O) e,ee, e,eae,,\x\<Vbe„ 21) Q{e,)) e,ee, e,ee,,\M<^<^i assumed n/^/n 0, ei+x/Vbse)" where Op(l), e,ee, Bee]" (13 = aee, Hence, by Claim 9/ C Cn{k) C 0;" (D.20) for +Op(l)), 1 dee)" ©/" (1 follows. proves Claim III 1 » dH{Cnik),e!) < dH{Cn(k),e,) < (D,18) 0/). + Op(l) - sup 4(5/, 0) 8;ee, 4(6'/,0) e,ee, from the uniform stochastic equicontinuity of the process A i--> sup^^ggg^ ib{Si,^) over K C.2 and from V^En^O, (D.22) which follows from the assumption that h/n (D.23) (b) follows (D24) —0 at polynomial rate, so that vftcn ex \/b/n by the Continuous Mapping Theorem and proof of Theorem sup 4(6/, ^e,ee, f 0), sup B,ede, 4(6'/,0))^d ' sup f ^ eiede,,>.eA^{e,) 3.1 In n — » 0; which implies that as ^oo(6'/,A), sup e,ede, 6 — oo t^[e,,0)\. ' D 35 Appendix Verification of A.l Proof of Theorem E. 3.3 immediate from the stated assumptions. Therefore we is will focus on verifying C.l to C.3. Step x'j6 d = > dim(6»). and < x'^O 96/ < Denote Tk := {Tk{xj),j T={tGR'^:71<f< is 9/ identified set T2 {xj) for all j < J}. J) for = A; determined by a set of inequalities: 0/ := [9 ^ is assumed that the d x J matrix It is = {eeR'^ .X'8 = T2} is compact and X X < := {xj,j J) is R'' some for t has t : rank, full < Ti 6/ < t T2}. compact. is assumed that ©/ C O. It is determined as 59/ = (E.2) That from any Of = G In{S) {61 G 9/ G int(9/) [Oi : : = x'^O infgjgae/ j5pr3^^«>o Suppose otherwise that k = \\Tk{xj)-x^9,\\ (E.4) ^ _ ^^1^, = T2 (xj) ^ ^/V^} ^'jW some for , j}. implies that x'^Oj = = x'.O'j Tk{xj) for > such that v{j,k). some j and k and some 9'j G 99/. This Hence 0. 99/ ^^^ ^ dQjy9'j G ^^11 must be bounded away observe that there exists k ||xj||5/y'n. Indeed, ^ei^dej,Wjede!:x'/j=Mxj), ^ dQj. Hence k > poses a contradiction to 9; ~ 11^/ This implies Xj9j 0. or x'j9 ti {xj) by the distance proportional to Tk (xj) (E-3) : x'/j = Tk{xj), \/{j, k). \\Xj\\ Recall that > ||xj|| for all j since the first 1 component \\Tk{xj)-x'Jj\\>K (E.5) of Xj is 1. V5/^96/, ||e^-^/||||xj|| inf Hence y{j,k). Thus, (E.6) 9i G In{5) => Ti [xj) - x'^9i < —^\\xj\\ or T2 (xj) - x'^Oi > Vj. -^||a;j||, Recall (E.7) 4 (e/, A) = ^ ^ fv^ j=i Iv (fi {x^) - J 12 x'^9,) - x'^x] L J -j- + j] !^ [^ (f2 [xi) - x'^9j) - x'^X\ L 11, J j=i Hence 2 (E.8) 4(e/,0) = ^^[w'i,-v/^(a:;e/-ri(a;,))l + j=i ^ 2 !^ [l^^^ - ^^ (x^^z - T2 (x,)) 1 . j=i Hence sup (E.9) 14 (0/, 0) I < e,ei^(5) Choose any (E.IO) : of rank Thus 1, 2. Bi (E.l) Since The Verification of C.l: I. Ti [xj) e > 0. Since ||xj|| limsupmaxP n—'oo ^ ^" \W,^ - k5\\x^ > 1, (Vl^ij Wij > = Op K(5||xj||) 111 ' -'+ Jit and W2J (1) < e and = Op 36 E- Jit (1) for " n—>c» [^2,- each j maxlimsupP J<J iSJ + (1^2^ + ^5\\x^ < J, for 5 sufficiently large, 111 ' < — '«<5||a:j||) < e- C.l is now verified, since it follows that for any sup (E.ll) e > there exists 5 sufficiently large such that 0, |4{6'/,0)| = wp > 1-e. e,ei„(6) Step II. Verification of part ft first that for fixed 61 G L 90/ and J 4- Mapping Theorem the J e^ (E,14) C 0/ , ^ 2 = Y.Pi Ki> - ^^^] 1 (^J^^ W' . C.2-ii(a) requires us to = "^1 (^j)) re < such that x'jfii (xji, Z = 1, ..., rank d and d) has ^ 90/, that = {kt,ji) is _ 1 [x'^di = the interior of T2 (Xj)) 0/ is . nonempty among The the finite subset of points systems of d-equations of the form: all Tki{xj,), a^^A] limit theory of supg^gge^ £„ {6j, ). 00 amounts to choosing Oj that are defined as a collection of solutions to (E.15) 0/ given by is 2 + Ep> K^ - examine the finite-dimensional of finding supg^ggg^ in (^/, A) for I, finite-dimensional limit of Cn {di A) Verification of part ii(a) of C.2: First, suppose that III. relative to Vj (01, X) Step in = 0wp ^1. 42'(9,,A) Therefore, by the Continuous problem by an argument similar to that fixed A, (E.13) Step where , -; (E.12) Note 4 (^/, A) = d^' [61, \) + d^' (5;, A) of C.2: Write i l € {1,2} x = l,...,d, {1, 2, ..., J} for each /. There is only a finite set of such systems of equations. Then, sup (E.16) e,ede, = max 4 (0/, A) ^'^v, 4 (61/, A) , for n < Hence by the Continuous Mapping Theorem the finite-dimensional weak by maxejgv, supe.eae, ^00 {^/i Oi s-nd therefore the finite-dimensional weak 00. maxe,gv, limit of 4 {^I^ ) limit of supg^gge^ £„ (0/, ) is is given given by ^oo(6'/,-)- 0/ = 90/, Second, suppose that the interior of 0/ empty relative to V =V 1 (n [x,] = T2 (.T,)) ^ (w,, - K'^. Then wp — » 1 ' x; a) '--^ (E.17) inf '^ Therefore, by Continuous infgjge, ^00 (^/i is 4 (^/, A) = inf ^ that ) The i^ ' (Or, A) 1 (ri (x,) = r^ {xj))pj [W.j - x'^xf . j<J Mapping Theorem the finite-dimensional limit of infg^ge; joint finite-dimensional convergence of (infe^ge, 37 4 (di, •) given by 4 (^7, 4 {di< )) follows iSupg^gQ, •) is by the Continuous Mapping Theorem. Step III. Verification of part ii(b) of C.2: Let fixed 6{ e 90/, in {^ii A) is K be any compact subset of R''. € K, where Vj tinuity for maxe^gv; ^n {di, A) over X a is finite subset of To dQj, it verify the stochastic equiconsuffices to show that for any € K. This immediately follows from convexity of stochastically equicontinuous in X ^n (^/! ) in A and finite-dimensional convergence oi £„ {9j, ) to a convex continuous function i^o (^/, ) over R"*, proven in Step I. Likewise, stochastic equicontinuity of infe,ge, in (^/, A) in A G /f also follows from convexity and finite-dimensional convergence of infe^ge; in to a convex continuous function infgjge/ ^oo (^/, •) (^/i •)' cf. (E,17). Step IV. Verification of part whenever ti [xj] = Xjdj, iii and Any A e of C.2: <0 x'jX whenever 56/ A<„ {d[) for 9j e = t2 [xj) satisfies the property that x'^X > Hence x'^Oj. J i^ sup (E.18) Vpj < {9j, A) + {[Wyj]l = Step V. Verification of C.3: For each 9n{di,X) such that u(9i,X) < [W2,]'i) infgjge, oo a.s. \\0i + X/^/n - > K/y/n, 9',\\ decompose (E.19) Any 61/ -I- A/x/n solution ^J is = e; -I- subject to A7v/n, where 1 <p< depends on ' {xj^,l = l,...,p) arg inf 8^dQ xr^9} = has rank p and {9;, A)) is necessarily = one of the Tk,{xj,), (fc;,j;) finitely r= (Tfc, are md 9\\ WX']] = y/nv {e,,X) > K. = l,...,p, {1,..., J} for each sub-matrices of I. Matrix X*'X* (which X*'X* that have full rank zero. Let also = ({ji,k,),l = l,...,p). zero, ci||A'||<||^*'A-|U. Moreover, given any index set JfC* for , x^,A'<0 (E.23) (E.22) at least for one {]' ,k') € (E.24) - X/y/n e {1,2} x bounded away from (E.22) By l {Xj,),l=l,...,p) and JIC* Since the eigenvalues of (^X*'X*) + many p x p and whose eigenvalues are bounded above away from (E.21) \\9i I dim(5) binding constraints of the form: (E.20) where X* := 6i; ci||A*|| < li^.A'l all [ji, ki) if fc, = 1 G JK.*: or x'^^X* >Q if fc; = 2. JK*: sothati^.A* < -Ci||A*|| 38 if fc* = 1 or x^. A' > ci||A*|| if k* = 2. p, = 4 (01, A) Decompose + d^' [9,, A) d^' (9/, A) where , (E.25) ^^ = 4^) (0;, A) >— - [l^ij- >p,.(l + (x^A*) 1' - 7^ [u/i, %/" (a^j- A*) ]' + o,{l))[Wrj. 1 =n (r;e; 1 (/=* = c,\\y\\]\ + 5 [vi/2, 1) + [n/^ (x;.A*) [k' = 1) >mmpj{l + 0p{l))\mm[Wij,-W2jJ < j<j ix,)) + - J] +Ci||A*||l J 1^2,.] ' 1 (fc* - W^j.]\ [cillA'll I - v^ {xry) [k' = 1 ' i (x^.e; = 2) = 2) r^ (x,)) , + and '-' (E.26) J J Hence, recalling that > constants c |H£| = wp > and Op{l), for any 1 — e > 0, 0, [m^2j - V^(x^r - = ^/nv{9!,\), one has that where W_ := min we can K select - X;A*]_l(x;e; < T2(X,)) > 0. — W2:;,i < W^ij, large (^/,A) (.„ j\- > c (H£ + c'||A*||)_^, Let \/nv{9j,X} = some for ||A*|| enough so that W_ > —c'y/nv(9j,X)/2 wp > > 1 positive ii". — e. Since Then e, el'\9j,\)>^cc'v{9,,\f. (E.27) Since Cn{9i,X) > 4^' (61/, A), C.3 is I. Verification of A. 1 is D verified. Appendix Step T2(X,)) =l ||A*|| > c' 2 E^ + F. Proof of Theorem 3.4 immediate from the stated assumptions and from {mi{9),9 G 8} being Donsker (hence Glivenko-Cantelli). Step II. Step III. Verification of C.l is not needed because 9/ 4(^7, A) (G„mi(0/ X Wn{9i condition ii. (F.2) condition (F.3) that {mi{9), GnTni{9i where A(9/) in 9. = v^E„mi(e/ + X/V^yWn{9i + X/^)V^Enmi{9, + = (F.l) By = 90/ Verification of C.2 Write is + + X^^/n) + VnEmi{9i + X/^/n) {Gnmi{9, e 9} forms a Donsker + X/^/n) class for + A/v^) = Gnmi{9,) + X/^/n) X/-Jn))' + y/^Em,{9i any compact Op(l) => A(0,), in + set X/^/n)) . K, L°°(9/ x /f) the Gaussian process with the covariance functions given in the statement of iv., Wn{9, + A/n/^) = W{9,) + 39 Op(l), in L°°(9/ x /C). Theorem 2.4. By By condition iii., ^Emi{ei + (F.4) = \l\fn) + G(e,)'X o(l), in L'^{Qi x K). Hence U9i, (F.5) Hence A) ^ ^oo(0/, A) = (A(0;) + G{e,y\)' C.2-i. is verified, since finite-dimensional C.2-ii.(a) and Un{X) = is is G(0/)'A) , in and Un(A) implied by weak convergence of is L~(e/ x K). implied by weak convergence in L°°{Qi x K). supg^gQ^ ^n(^/i A) are continuous transformations of £n(0, stochastic equicontinuity of ^^(A) which convergence + by the Continuous Mapping Theorem since the functionals Zn(A) verified is x W^(e,) x (A(0;) = inie,£e, ini^i,^) C.2-ii.(b) follows as well, since A). implied by stochastic equicontinuity of £n(^/,A) over 6/ x K, (-niSi, A) in L°°(0/ x K) to the quadratic form of a Gaussian process, ^oc(0/,A). Step IV. Verification of C.3. By condition ii. G,m.(e) (F.6) where A(0) mi{9) forms a Donsker ^ A(e) Hence class. L°°(G), in the Gaussian process with the covariance functions given in the statement of is Theorem 2.4. Hence (F.7) = sup||G„mi(5')|| 0p(l). see By condition iv., iv. fn -^p > ^ uniformly inO € Hence wp 0. — > Q Wn{d) > i ( > (F.8) inf in inf ^,n\\ v(B,,x)>K/V^ ~ Note the By line different is Define ^n mineig {Wn{9)). infflge By condition viB,,x)>K/V^ from the preceding one, as it ' VnEmi{ei + X/^) Op{l) II y/n(v{9i,X)^A6'^) ^n{v{9!,XyA5^)\\ II y/n{v{9i,xyA5^) ^7i{v{9j,xyAS^)\\oo uses the sup norm instead of the Euclidean norm. the partial identification condition (3.15) |[v/^£m.(e;+A/0i)|| ^ (F,9) II for = (^W^Emiid, + \/y^) + Gnm^iOi + A/v^)||2) USi, ( ;;"::,^! „^ . /^\n{v{e[,X) inf — W{9) + 0p{l). 1 some constant C> 0. (F.IO) For a given e > Hoc y/n{v{ei,X)^A6^) 0, II— K can be made arbitrarily large, so that =£diL_|| wp > 1 — e <C/2. IVn(i)(6'/,A)2A<52)lloo Hence wp > 1 — e, v{B,,x)>K/y/^\n{v{9i,Xy A5^)J Hence wp > 1 — Hence C.3 is v / / e, en{9!,X)>C' n-{v{9i,Xf (F.12) •>/ AS'^) for all i/(6i/, A) > J-C/v^. D verified. 40 Appendix Suppose that one when well motivated, when analysis or some parameter interested in a special parameter 9* inside is there it is value, rV a sense in which 9* is 0/ The . 9* in some inference about 0/ is the "truth". This case typically arises in non-structural is believed that the models are correct representations of data-generating processes for there i.e. of data P. In this scenario, 0/ 9' is e such that the model law Pe agrees with the actual stochastic law R'' not of interest per is problem has been already investigated similar to Pointwise Inference G. Relationship quantile estimation by Chernozhukov and se, but rather 9' in the context of Hansen (2003), is. In method of moments settings, a dynamic model with censoring by Hu and (2002), GMM weak or complete unidentification by Kleibergen (2002). Imbens and Manski (2004) investigate the Wald type inference about 9* for the special case where a real parameter of interest is known to lie in an interval with endpoints that can be consistently Here we provide further insights concerning the pointwise inference estimated. in its relation to regionwise inference. Assumption A. 4. Suppose there exists On aniQniOl) where C{9i) a is random has continuous distribution function on G.l. Suppose liminf„_cx> P{e* e C„(c;)| > that > ~ oo such that Qn) -^d Moreover, for at variable. Theorem — (0, 0(9!) least for dl one 9j € 0/, C{9i) oo); otherwise, C{9j) Assumptions A.l and A. 4 0/6 0/ = > with positive probability and with probability one. hold. Let c* = supg^ Ca{9j). Then for any 9* in Qj a. Proof: liminf P{0* e CnK)} = liminf P{a„ (Q„(9') - qn) < supc„(6')} (G.l) (1) > liminfP{a„(Q„(e*)-g„) <c„(6'-)} > (1 or a) > a, D The theorem above is constructed using the Anderson-Rubin pointwise testing principle. Notice that the confidence set constructed in the theorem above Theorem C„(c„(-)) (G.2) also has the necessarily smaller than the regionwise confidence set in is Notice also that as a byproduct of our derivation, inequality 2.1. a pointwise coverage. Hu = {r e 0/ : On {Qn{9') - Qn) is generally not equal is generally a special case Qnid) (is let also use this technique for this set for the case of interval-identified parameter. ^^ of our construction. all The is quantiles. However, this set 9 € 0/.'" Ca{9) be the subsampling estimate of Ca{9) for each 9 e Qi. inequalities. '''This IV set Cn{ca{-)) This can be seen by defining the new objective function ^^A more recent paper than ours, by Ho, Pakes and Porter (2004), has also proposed such sets moment set < cU9')} smaller) than the level set C„{c*) of the function OnQniS)- = Q„{9)/m3x[cai9),e] for Further shows that the (2001) proposed the set (G.2) in the context of a partially identified dynamic censored regression model. Chernozhukov and Hansen (2003) Manski and Imbens (2004) construct (1) in (G.3) The difference is an equi-quantile transformation of the original objective function. In many examples as objective functions in the context of that they do not directly work with objective functions. have the equi-quantile property by using optimal weights. 41 this is unnecessary, Theorem G.2. Suppose that Assumptions A.l and A. 4 hold. Let liminf„_,oo -P] ^* Proof: € Cr,{c*^) > 0/ C 9; wp Since liminf >a- — > P{r (Likewise, a corollary 1, it e C„(c;)) = liminf n—too 9* in Q; supc<,(^)} Qn) < F{a„ (Qn(r) - g„) < c,(r) + > liminf (1 < Qn) - or q) for > Ccie*)} Op(l)} a, subsampling, Inequality (1) shows that C„(c* ( Then for any q "^°° from the standard argument 2.1. P{a„ {Qn{e*) - limini PianiQuiO") > (2) follows sup^^ Ca{9[). that Cn{c*^{-)) also covers 6' with probability a.) is > (G.3) Equality = follows that n—*oo proof of Theorem c* as the one presented in Step 2 of the e.g. also covers 6* with probability a. (•)) ). D [CONSISTENCY AND RATES HERE TOO?] Notation and Terms —>p —d wp -^ wp > 1 — convergence in (outer) probability P* convergence in distribution under P* 1 with inner probability P, converging to one e with inner probability P, larger than closed ball centered at x of radius 5 Bs{x) I > 1 — e for sufficiently large n, Q identity inatrix normal random vector with mean A/'(0, a) T Donsker class here this and variance matrix a means that empirical process / >—» -4= 5I)"=i(/(^i) ~ EfiyVi)) is asymptotically Gaussian in L°°{J^), see Vaart (1999) metric space of bounded over L°°{J-) minimum mineig(A) T eigenvalue of matrix functions, see Vaart (1999) A References Andrews, D. W. K. (1999): "Estimation when a parameter "Inconsistency of the bootstrap (2000); is on a boundary," Econometrica, 67(6), 1341-1383. when a parameter is on the boundary of the parameter space," Econometrica, 68(2), 399-405. Angrist, J., AND Chernozhukov, v., Krueger (1992); "Age at School Entry," JASA, 57, 11-25. AND C. Hansen (2001); "An IV Model of Quantile Treatment v., and AND E. A. Effects," forthcoming in Econo- metrica. Chernozhukov, H. Hong (2003); "An MCMC Approach to Classical Estimation," Journal of Econometrics, 115, 293-346. CiLlBERTO, F., Tamer (2003); "Market Structure and Multiple Equilibria in Airline Markets," Working Paper: Princeton University. Demidenko, E. (2000): "Is this the least squares estimate?," Biometrika, 87(2), 437-452. GiSLTEiN, C. Z., Hansen, L., J. and E. Leamer Heaton, and of Finacial Studies, 8, (1983); "Robust Sets of Regression Estimates," Econometrica, 51. E. G. Luttmer (1995); "Econometric Evaluation of Asset Pricing Models," 237-274. 42 The Review Horowitz, and J., C. Manski "Identification (1995): and Robustness with Contaminated and Corrupted Data," Econometrica, 63, 281-302. "Censoring of Outcomes and Regressors Due to Survey Nonresponse: Identification and Estimation (1998): Using Weights and Imputations," Journal of Econometrics, 84, 37-58. Hu, L. (2002); "Estimation of a Censored AND IMBENS, G., Manski C. (2004): Dynamic Panel Data Model," Econometrica, 70(6). "Confidence Intervals for Partially Identified Parameters," Department of Economics, Northwestern University. Kleibergen, F. (2002): "Pivotal statistics for testing structural parameters in instrumental variables regression," Econometrica, 70(5), 1781-1803. Klepper, AND S., Leamer E. (1984): "Consistent Sets of Estimates for Regressions with Errors in All Variables," Econometrica, 52(1), 163-184. Manski, C. (2003): Partial Identification of Probability Distributions, Springer Series in Statistics. Springer, Tamer Manski, C, and E. "Inference on Regressions with Interval (2002): Data on a Regressor New York. or Outcome," Econometrica, 70, 519-547. Marschak, AND W. J., H. Andrews (1944): "Random Simultaneous Equations and the Theory of Production," Econometrica, 812, 143-203. MiLNOR, J. (1964): "Differential topology," in Lectures on Modern Mathematics, Vol. II, pp. 165-183. Wiley, New York. Nerlove, M. Estimation and Identification of Cobb-Douglas Production Functions. Rand McNally (1965): & and Company, Chicago. Newey, W. and K., D. McFadden (1994): "Large sample estimation and hypothesis testing," in Handbook of econometrics, Vol, IV, vol. 2 of Handbooks in Econom., pp. 2111-2245. North-Holland, Amsterdam. Robert, C. P., AND G. Casella (1998): Monte Carlo VAN DER Vaart, a. W., and New J. A. Wellner (1996): Statistical Methods. Springer-Verlag. Weak convergence and empirical processes. Springer-Verlag, York. POLITIS, D., J. Romano, and M. Wolf Vaart, A. V. D. (1999): Subsampling, Springer Series in Statistics. Springer:New York. (1999): Asymptotics Statistics. Cambridge University 43 ^2 Press. Date Due Lib-26-67