Comments on “Partial Identification by Extending Subdistributions” by Alexander Torgovitsky Frank A. Wolak Department of Economics Director, Program on Energy and Sustainable Development Stanford University Stanford, CA 94305-6072 wolak@zia.stanford.edu http://www.stanford.edu/~wolak Motivation for Research • Obtaining point identification of economic magnitude of interest often requires difficult-to-defend distributional assumptions or functional form assumptions on econometric model • Partial identification modeling framework provides alternative approach to estimating economic magnitude of interest without imposing these assumptions – Advantage—Researcher only imposes assumptions on distribution of unobservables and functional forms of econometric model that he/she finds credible – Disadvantages--Researcher can typically only estimate identified set that contains true economic magnitude of interest – Extremely challenging numerical problem to compute estimate of identified set – Computationally intensive procedures for testing hypotheses about characteristics of identified set or points in identified set Summary of Results • Main Result---Partial Identification by extending subdistributions (PIES) • General econometric model – Y = h(X,U) – Y = vector of outcome variables – X = vector of explanatory variables – U = L-dimensional vector of latent variables with conditional distribution U|X = x given by F(u|x) – h(u,x) = structural function • Researcher makes assumptions about F(u|x) and h(x,u) that identifies the set that contains these magnitudes from conditional distribution of Y given X – h and F that satisfy these assumptions are called admissible values • By definition, h is in the identified set if and only if there exists an admissible F that generates the observed distribution of Y given X Summary of Results • This requirement constrains the behavior F(u|x) on subset Ux(h) of π πΏ – Author calls restriction of F(u|x) to Ux(h) a subdistribution – Has properties of distribution function on Ux(h) • Main Theoretical Result --Fix admissible h, if there exists a subdistribution function, πΉ π’ π₯ , on Ux(h) for each x that satisfies observational equivalence condition, then subdistribution can be extended to a distribution function F(u|x) on π πΏ that yields observed distribution of Y|X=x • Importance of result is that restrictions that determine whether a function is a subdistribution are linear constraints on the values of the function πΉ π’ π₯ Summary of Results • Paper applies result to ordered discrete response model • π= π½ π=1 π¦π 1[ππ−1 π < π ≤ ππ π ] • • • • • {g0,g1, …, gJ} = a vector of functions X U = scalar latent variable {y1, …, yJ} = discrete support of Y {x1, …, xK} = discrete support of X Computes values of identified set for binary response model with g1(X) = β0 + β1X1 + β2X2 • Considers case that X1 exogenous and X1 is endogenous – X1 exogenous cases considered—(1) X and U are independent, (2) median of U given X is zero, (3) U given X is symmetric around zero – X1 endogenous, same cases considered as well additional cases for latent variable in second equation of model determining value of indicator Y2 (endogenous X1) that depends on instrument X3 Summary of Results • Extends PIES framework to compute identified set for average structural function (ASF) for binary response model and average treatment effect (ATE) – E(Y1 |X2,Y2 = 1) and E(Y1 |X2,Y2 = 0) – Average Treatment Effect is difference of ASFs – For some assumptions on binary choice model with endogenous right-hand side variable identified set for ATE is not connected • Results in Table 2 • PIES framework extended to derive subdistribution extension lemma for general modeling framework – PIES applied to two-sector Roy model in abstract form but no identified sets where computed – Applying procedure to compute identified sets for more general models likely to be challenging Comments on Paper • General comment on partial identification literature – Theory-based empirical researchers are very sympathetic to this approach, but it is hard to convince other empirical researchers of its merits given the lack of examples demonstrating empirical content • Are there simple examples to illustrate how to use estimation and inference procedures on an important empirical question? – Example that demonstrates that assumptions needed for point identification can yield estimates that are outside identified set for more credible assumptions • Can computer software or detailed instructions on how to implement procedures be provided for a class of empirical problems? Comments on Paper • Can realistic Monte Carlo studies be run illustrating – Biases in common parametric approaches that are not present in partial identification approaches – Credible identifying assumptions that can still yield informative answers about economic magnitudes interest from identified set • Identified set of demand price elasticity • Identified set of compensating variation associated with price change • Partial identification approach offers way for economic theory to be used to measure magnitudes of economic interest without “incredible assumptions” needed for point identification Comments on Paper • Specific comments/questions about paper • More details on procedure used to solve for identified sets would be very informative • More discussion of cause of results in Table 2 – Disconnected identified sets • More discussion of specific classes of models PIES approach could be applied to would be useful • More discussion of ways to relax linear functional form assumption on g(X) function – Particularly for binary response models, linear index seems more restrictive than distributional assumption on latent variable • Amemiya (1981) derives approximate relationships between probability limit of slope coefficients in linear index model for probit, logit and linear probability models Concluding Comments • Partial identification methods have potential to “put the economics back into econometrics” • To do so researchers must • Show practical usefulness of partial identification methods to empirical researchers • Illustrate relationship between assumptions researcher is willing to make and form of identified set for some commonly employed model • Provide software and more details on how to implement methods • Simple to implement rules-of-thumb may be preferable to rigorous, but difficult to implement approaches Thank You