The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Federal Reserve Bank of Chicago Estimation of Panel Data Regression Models with Two-Sided Censoring or Truncation Sule Alan, Bo E. Honoré, Luojia Hu, and Søren Leth–Petersen WP 2011-08 Estimation of Panel Data Regression Models with Two-Sided Censoring or Truncation Sule Alany Bo E. Honoréz Luojia Hu x Søren Leth– Petersen { November 14, 2011 Abstract This paper constructs estimators for panel data regression models with individual speci…c heterogeneity and two– sided censoring and truncation. Following Powell (1986) the estimation strategy is based on moment conditions constructed from re– censored or re– truncated residuals. While these moment conditions do not identify the parameter of interest, they can be used to motivate objective functions that do. We apply one of the estimators to study the e¤ect of a Danish tax reform on household portfolio choice. The idea behind the estimators can also be used in a cross sectional setting. Key Words: Panel Data, Censored Regression, Truncated Regression. JEL Code: C20, C23, C24. This research was supported by NSF Grant No. SES-0417895 to Princeton University, the Gregory C. Chow Econometric Research Program at Princeton University, and the Danish National Research Foundation, through CAM at the University of Copenhagen (Honoré) and the Danish Social Science Research Council (Leth– Petersen). We thank Christian Scheuer and numerous seminar participants for helpful comments. The opinions expressed here are those of the authors and not necessarily those of the Federal Reserve Bank of Chicago or the Federal Reserve System. y Faculty of Economics, University of Cambridge, Sidgwick Avenue, Cambridge, UK, CB3 9DD. Email: sule.alan@econ.cam.ac.uk. z Department of Economics, Princeton University, Princeton, NJ 08544-1021. Email: honore@Princeton.edu. x Economic Research Department, Federal Reserve Bank of Chicago, 230 S. La Salle Street, Chicago, IL 60604. Email: lhu@frbchi.org. { Department of Economics, University of Copenhagen, Øster Farimagsgade 5, Building 26, DK-1353 Copenhagen K. and SFI, The Danish National Centre for Social Research, Herluf Trolles Gade 11, DK-1052. Email : soren.lethpetersen@econ.ku.dk. 1 1 Introduction This paper generalizes a class of estimators for truncated and censored regression models to allow for two– sided truncation or censoring. The class of estimators is based on pairwise comparisons and the proposed generalizations therefore apply to both panel data and cross sectional data. A leading example of when two– sided censored regression models are useful is when the dependent variable is a fraction. For example, Alan and Leth-Petersen (2006) estimate a portfolio share equation where the portfolio shares are between 0 and 1, with a signi…cant number of observations on either of the limits. Other recent applications in economics of regression models with two– sided censoring include Lafontaine (1993), Petersen and Rajan (1994), Petersen and Rajan (1995), Houston and Ryngaert (1997), Fehr, Kirchler, Weichbold, and Gachter (1998), Huang and Hauser (1998), McMillan and Woodru¤ (1999), de Figueriredo and Tiller (2001), Huang and Hauser (2001), Fenn and Liang (2001), Poterba and Samwick (2002), Nickerson and Silverman (2003), Of…cer (2004), Charness, Frechette, and Kagel (2004), Andrews, Schank, and Simmons (2005) and Gi¤ord and Bernard (2005). We formulate the two– sided censored regression model as observations of (y; x; L; U ) from the model y = x0 + " where y is unobserved, but we observe 8 > L > > < y= y > > > : U and if (1) y <L if L if y (2) U y >U is the parameter of interest. When y is a share, L and U will typically be 0 and 1, respectively. We will focus on panel data settings so the observations are indexed by i and t where i = 1; : : : ; n and t = 1; : : : Ti . This allows for unbalanced panels, but we will maintain the restrictive assumption that Ti is exogenous in the sense that it satis…es all the assumptions made on the explanatory variables. In a panel data setting, it is also important to allow for individual speci…c e¤ects in the errors "it . We will do this implicitly by making assumptions of the type that "it is stationary conditional on (xi1 : : : xiTi ) or that "it and "is are independent and identically distributed conditional on some unobserved component ist . These will have the textbook speci…cation "it = vi + it (f it g i.i.d.) as a special case. In section 4.2, we discuss how to apply the same ideas to construct estimators of the cross sectional version of the model. For the two– sided truncated regression model, we assume observations of (y; x; L; U ) from the 2 distribution of (y ; x; L; U ) conditional on L y U . Two– sided truncated regression models are less common than two– sided censored regression models, but they play a role in duration models. Suppose, for example, that one wants to study the e¤ect of early life circumstances on longevity by linking the Social Security Administration’ Death Master File to the 1900– s 1930 U.S. censuses1 . This will miss a substantial number of deaths: (1) individuals who died under age 65 (since they were less likely to be collecting social security bene…ts and thus their deaths were less likely to be captured); (2) individuals who died before the year 1965 (the beginning of the computerized Social Security …les); and (3) individuals who died after the last year the data is available. This is a case of two-sided truncation because we observe an individual only if the dependent variable, age at death, is greater than 65 years and if the death occurs between 1965 and the last year the data is available. This is essentially the empirical setting in Ferrie and Rolf (2011), although they only consider data from the 1900 census, so right– truncation is unlikely to be an issue in their application. Honoré (1992) constructed moment conditions for similar panel data models with one– sided truncation or censoring and showed how they can be interpreted as the …rst– order conditions for a population minimization problem that uniquely identi…es the parameter vector, . This paper generalizes that approach to the case when the truncation or censoring is two– sided. The main contribution of the paper is to show that some of those moment conditions can be turned into a minimization problem that actually uniquely identi…es when there is two– sided censoring or truncation. This is an important step because the moment conditions that we derive do not identify the parameters of the model. This is a generic problem with constructing estimators based on moment conditions. For example, Powell (1986) constructed moment conditions for a related cross sectional truncated and censored regression models based on symmetry of the error distribution. He also pointed out that while these moment conditions did not identify the parameter of interest, minimization of an objective function based on them did lead to identi…cation. The rest of the paper is organized as follows. Section 2 derives the moment conditions and the associated objective function for models with two– sided censoring and two– sided truncation. Section 3 then discusses how these can be used to estimate the parameters of interest. Generalizations of the two models considered in Section 2 are discussed in Section 4. We present the empirical application in Section 5 and Section 6 concludes. 1 1930 is the latest year for which this linkage is feasible. 3 2 Identi…cation: Moment Conditions and Objective Functions The challenge in constructing moment conditions in models with censoring and truncation is that one typically starts with assumptions on " conditional on x in (1). If one had a random sample of (y ; x) then these assumptions could be used immediately to construct moment conditions. For example if E ["x] = 0, then one has the moment conditions E [(y x0 ) x] = 0. However with truncation or censoring, y x0 will not have the same properties as ". The idea employed in Powell (1986), Honoré (1992) and Honoré and Powell (1994) is to apply additional censoring and truncation to y x0 in such a way that the the resulting re– censored or re– truncated residual satis…es the conditions assumed on ". For example, Powell (1986) assumed that " is symmetric conditional on x in a censored regression model with censoring from below at 0. If that is the case, then y x0 = max f"; x0 g will clearly not be symmetric conditional on x, but the re– censored residuals, min fy x0 ; x0 g will be. This implies moment conditions of the type E [min fy Unfortunately, this moment condition will not in general identify to prove that the integral (as a function of b) of E [min fy b = x0 ; x0 g x] = 0. , but Powell (1986) was able x0 b; x0 bg x] is uniquely minimized at under appropriate regularity assumptions. Honoré (1992) applied the same insight to a panel setting where the assumption was that " is stationary conditional on the entire sequence of explanatory variables. Again, censoring or truncation destroys this stationarity, but it can be restored for a pair of residuals by additional censoring. Honoré and Powell (1994) then applied the same idea to any pair of observations in a cross section, and Hu (2002) generalized it to allow for lagged latent dependent variables as covariates. The contribution of this paper is to generalize the approach in Honoré (1992) to the case with two– sided censoring or truncation. As in Powell (1986), Honoré (1992) and Honoré and Powell (1994), it is straightforward to construct moment conditions based on some re– censored or re– truncated residuals. However, it is not clear that these moment conditions will identify the parameters of interest, and we therefore construct (population) objective functions from these moment conditions, and then explicitly verify that these objective functions are uniquely minimized at the parameter. It is the construction of the objective functions and verifying that they are uniquely minimized at the true parameter value that constitute the methodological contribution of the paper. The general approach is to start with a comparison of two observations for a given individual in a panel. Based on these observations we will construct re– censored or re– truncated residuals 4 eits (yit ; xit ; xis ; Lit ; Lis ; Uit ; Uis ; b) and eist (yis ; xis ; xit ; Lit ; Lis ; Uit ; Uis ; b) that have the same properties as "it and "is when b = . This will then imply that if "it and "is are identically distributed conditional on (xit ; xis ), then E [ (eits (yit ; xit ; xis ; Lit ; Lis ; Uit ; Uis ; ) eist (yis ; xis ; xit ; Lit ; Lis ; Uit ; Uis ; ))j xit ; xis ] = 0: (3) This will form the basis for construction of our estimators. Of course, once it has been established that eits (yit ; xit ; xis ; Lit ; Lis ; Uit ; Uis ; ) and eist (yis ; xis ; xit ; Lit ; Lis ; Uit ; Uis ; ) are identically distributed then for any function, ( ), we also have the moment condition E [ ( (eits (yit ; xit ; xis ; Lit ; Lis ; Uit ; Uis ; )) (eist (yis ; xis ; xit ; Lit ; Lis ; Uit ; Uis ; )))j xit ; xis ] = 0: (4) provided that the moment exists. Moreover, if the errors "it and "is are also independent conditional on (xit ; xis ), then so are eits (yit ; xit ; xis ; Lit ; Lis ; Uit ; Uis ; ) and eist (yis ; xis ; xit ; Lit ; Lis ; Uit ; Uis ; ). This implies that their di¤erence is symmetrically distributed around 0, so for any odd function ( ), E [ ( (eits (yit ; xit ; xis ; Lit ; Lis ; Uit ; Uis ; ) eist (yis ; xis ; xit ; Lit ; Lis ; Uit ; Uis ; )))j xit ; xis ] = 0: (5) provided that the moment exists.2 In this paper we will focus on (3) and the generalization (5). The reason for this is that in a linear model without censoring or truncation, (3) will correspond to OLS on the di¤erenced data, whereas (5) will also accommodate least absolute deviation estimation on the di¤erenced data as a special case. As already mentioned, the construction of the residuals eits (yit ; xit ; xis ; Lit ; Lis ; Uit ; Uis ; ) is fairly straightforward, and the challenge is to show that although (3) and (5) may not identify , and unconditional version of them can be integrated to yield a population objective function that is uniquely minimized at b = . This will involve the integral of ( ), which we will denote by ( ), and in addition to being odd, we will assume that ( ) is also increasing, so symmetric function. The leading cases are (d) = jdj and ( ) is a convex (d) = d2 . In order to simplify the exposition, we will …rst develop the case when Lit = 0 and Uit = 1. We will then demonstrate that the result can be adapted to the general case. 2 Of course one could combine the insight in (4) and (5) to get even more general moment conditions. See also the discussion in Arellano and Honoré (2001) and Honoré and Hu (2004). 5 2.1 Two– Sided Censoring Consider …rst the situation with two– sided censoring. Consider an individual, i, in two time periods, t and s, and assume that "it and "is are identically distributed. The distribution of yit be the same as that of "it except that the former is censored from below at above at 1 x0 it will and from x0 . Figure 1 illustrates this. The dotted line depicts the distribution of "it , while it x0 , which typically has point mass at it the solid line gives the distribution of yit 1 x0 it x0 it (illustrated by the fatter vertical lines). Since x0 it x0 it distributions of yit x0 is and yis x0 it and will typically di¤er from x0 , the is (given (xit ; xis )) will di¤er even if f"it g is stationary (given (xit ; xis )). However, it is clear that one could obtain identically distributed “residuals” by arti…cially censoring yit min f1 x0 ; 1 it x0 and yis it x0 from below at max f x0 ; x0 g and from above at is it is x0 g. See the dashed lines in Figure 1. One can then form moment conditions is from the fact that the di¤erence in these “re– censored”residuals will be orthogonal to functions of (xit ; xis ).3 Of course, this construction is only useful if the supports of yit x0 and yis it 1 < x0 it x0 < 1, because otherwise, is x0 will not overlap. is In order to proceed, we need explicit expressions for the di¤erence in these “re– censored”residuals. Consider …rst the case when x0 it x0 . Then the di¤erence in the arti…cially censored is residuals for individual i in periods t and s is max yit = max yit = max yit and when x0 it x0 it x0 it x0 ; 1 it min yit ; 1 + x0 it = min yit ; 1 + x0 it 3 x0 ; 1 is min yis x0 is x0 is ;0 ;0 x0 is x0 it (6) x0 it min yis ; 1 min yis ; 1 x0 it x0 is x0 is x0 is ; x0 is min yit = x0 ; x0 it is x0 is x0 is x0 is max yis x0 it x0 ; x0 is it (7) max yis + x0 it max yis + x0 it x0 is x0 is ;0 x0 it ;0 : Clearly, one can also use the fact that di¤erences in functions of the re– censored residuals will be orthogonal to functions for the explanatory variables. As discussed in Arellano and Honoré (2001), one can also construct moment conditions based on symmetry under the additional assumption that ("i1 ; :::; "iT ) is exchangeable conditional on (xi1 ; :::; xiT ). This is the motivation for the approach in Honoré (1992). 6 Figure 1: Illustration of Re– Censored Residuals when x0 > x0 . it is If we de…ne 8 > 0 > > > > > > 1+d > > > > > > min f1 y2 ; y1 g > > < u (y1 ; y2 ; d) = y1 y2 d > > > > max fy > 1; y2 g 1 > > > > > > d 1 > > > > : 0 where c1 = min f y2 ; y1 then u (yit ; yis ; x0 it 1g, c2 = max f y2 ; y1 for d< 1 for 1 < d < c1 for c1 < d < c2 for c2 < d < c3 for c3 < d < c4 for c4 < d < 1 for d>1 1g, c3 = min f1 y2 ; y1 g and c4 = max f1 (8) y2 ; y 1 g x0 ) will give the di¤erence in the re– censored residuals discussed above (see is Appendix 1). Hence the moment conditions are E u yit ; yis ; x0 it x0 is xit ; xis = 0; which implies the unconditional moments E u yit ; yis ; (xit xis )0 (xit xis ) = 0: (9) Panel A of Figures 2– depict the contribution to the moment condition function u (y1 ; y2 ; d) 4 for pairs of uncensored observations (Figure 2), 7 pairs with one censored and one uncensored Figure 2: The Functions u (y1 ; y2 ; ) and U (y1 ; y2 ; ). Neither Observation Censored. observation (Figure 3) and pairs with one observation censored from above and one from below (Figure 4). Although the true parameter value will satisfy (9), it is not in general the unique solution to the moment condition. This is illustrated in Panel A of Figure 5. It considers the case when yi1 N (0:5; 1) and yi2 N (0:4; 1) and both are censored from below at 0 and from above at 1. Note that this is the data generation process that one would get with = 0:1, xi1 = 1 and xi2 = 0 for all i. It is clear from Figure 5 that the moment condition E [ u (yit ; yis ; x0 b it x0 b)j xit ; xis ] = 0 does is not identify the parameter, , of the model. The most obvious reason is that, as mentioned, only observations for which 1 < (xi1 xi2 )0 b < 1 will contribute. In this case xi1 xi2 = 1 for all observations, so the moment condition is automatically satis…ed when jbj > 1. Following Powell (1986), we attempt to overcome the non– identi…cation based on the moment condition by turning it into the …rst order condition for a minimization problem. It is easy to see that (9) is (half of minus) the …rst order condition for minimizing E U yit ; yis ; (xit 8 xis )0 b ; (10) Figure 3: The Functions u (y1 ; y2 ; ) and U (y1 ; y2 ; ). One Observation Censored. Figure 4: The Functions u (y1 ; y2 ; ) and U (y1 ; y2 ; ). Both Observations Censored. 9 Figure 5: The Functions E [u (y1 ; y2 ; )] and E [U (y1 ; y2 ; )]. where U (y1 ; y2 ; d) = 8 > > > > > > > > > > > > > > > < > > > > > > > > > > > > > > > : 1 + 2c1 + c2 1 2d 2c3 c1 + 2c3 c2 + (y1 d2 + 2c1 + c2 1 2c3 c1 + 2c3 c2 + (y1 2c3 d + 2c3 c2 + (y1 (y1 y2 1 + c2 4 2c4 2c4 c2 )2 for 2 c2 ) d< 1 y2 2c2 c4 + 2c2 c3 + (y1 2c2 c4 + 2c2 c3 + (y1 y2 2 y2 c3 ) 2 c3 ) c1 < d < c2 c2 < d < c3 : for c3 )2 1 < d < c1 for c2 ) 2 for for y2 y2 2 d) 2c2 d + 2c2 c3 + (y1 d2 + 2d + c2 4 y2 c3 < d < c4 for c4 < d < 1 for d>1 Panel B of Figures 2– depict the contribution to the objective function U (y1 ; y2 ; d) for pairs 4 of uncensored observations (Figure 2), pairs with one censored and one uncensored observation (Figure 3) and pairs with one observation censored from above and one from below (Figure 4). Like the estimator for the panel data one– sided censored regression model developed in Honoré (1992), the objective function is piecewise quadratic or linear. However, surprisingly, going from one– sided to two– sided censoring ruins the convexity of the objective function, and the shape of the function is more similar to the objective function in Powell (1986) although that estimator was developed for a cross section model with symmetrically distributed errors. It is clear from the discussion above that the true 10 will solve the …rst order condition for minimizing the population objective function in (10). However, since U is constant, linear, quadratic and convex, and quadratic and concave over di¤erent regions, it is not at all obvious that will be the unique solution to these …rst order conditions. The key step for establishing identi…cation of is therefore to establish that the function in (10) is minimized at . We establish this in Appendix 1 (Section 7.1), and the result is illustrated in Panel B of Figure 5. As mentioned, it is also possible to construct moment conditions based on (5). Let and symmetric, and let () = 0( be convex ) (when it exists). When "it and "is are independent and identically distributed, we also have the moment conditions E u yit ; yis ; x0 it x0 is xit ; xis = 0 which imply the unconditional moments E where (u (y1 ; y2 ; d)) = u yit ; yis ; (xit 8 > > > > > > > > > > > > > > > < > > > > > > > > > > > > > > > : 0 xis )0 (xit for xis ) = 0 d< (11) 1 (1 + d) for 1 < d < min f y2 ; y1 1g = c1 (c3 ) for c1 < d < max f y2 ; y1 1g = c2 y2 d) for c2 < d < min f1 y2 ; y1 g = c3 (c2 ) for c3 < d < max f1 y2 ; y1 g = c4 (y1 (d 1) for for 0 c4 < d < 1 d>1 Except for a multiplicative constant, (11) is the …rst order condition for minimizing E U yit ; yis ; (xit xis )0 b (12) where U is found by integrating (u (y1 ; y2 ; d)) over each of the regions and insisting on continuity at the boundaries between the regions: 8 > (0) (1 + c1 ) > > > > > > (1 + d) (1 + c1 ) > > > > > > > (c3 ) d + > < U (y1 ; y2 ; d) = > > > > > (c2 ) d + > > > > > (d 1) > (c4 1) > > > > : (0) (c4 1) (c3 ) c1 + (c3 ) c2 + (c3 ) c1 + (c3 ) c2 + (c3 ) c2 + (y1 (y1 y2 (y1 y2 y2 y2 (c2 ) c4 + (c2 ) c3 + (c2 ) c4 + (c2 ) c3 + d< 1 c2 ) for 1 < d < c1 for c1 < d < c2 c2 < d < c3 : for c3 < d < c4 c3 ) for c4 < d < 1 c2 ) c3 ) (y1 (y1 y2 y2 The appendix establishes that (12) is uniquely minimized at b = . 11 for for y2 (y1 c2 ) d) (c2 ) c3 + (y1 c3 ) for d>1 2.2 Two– sided Truncation Mimicing the argument for the model with two– sided censoring, it is clear that if "it and "is are independent and identically distributed conditional on (xit ; xis ), then the observed errors, yit x0 , it will be i.i.d. except that the sampling scheme will have truncated them at di¤erent points, and x0 is x0 it from below and 1 x0 is and 1 from above. We can then construct identically x0 at it distributed residuals by arti…cially truncating yit above (and similarly for yis E r yit ; yis ; (xit x0 it x0 from below and at 1 is x0 from is x0 ). This yields many moment conditions, including is xis )0 E yit yis x0 it 1 xis )0 (xit 1 x0 is yis x0 it 1 x0 is x0 it yit (xit 1 xis ) x0 is (13) = 0 It is an easy exercise to see that except for a multiplicative constant (13) is the …rst order condition for minimizing xis )0 b E R yit ; yis ; (xit where R (y1 ; y2 ; d) is de…ned by 8 > 1 (y1 y2 max fy1 > 2 > < > > > : 1 2 1 2 (y1 (y1 y2 y2 1; y2 g)2 if d)2 min fy1 ; 1 y2 g)2 (14) max fy1 1; y2 g > d if max fy1 1; y2 g d min fy1 ; 1 if d > min fy1 ; 1 y2 g y2 g Figure 6 depicts the function R and its derivative, whereas Figure 7 shows their expectation when y1 N (0:5; 1) and y2 N (0:4; 1) and both are truncated from below at 0 and from above at 1. More generally, again let be convex and symmetric, and let () = 0( ) (when it exists). Then xis )0 yit yis (xit 1 x0 is yit x0 it 1 1 E x0 it x0 is yis x0 is 1 x0 it (xit xis ) = 0 1 (xit xis ) = 0 or E yit yis (xit xis )0 1 0 yit xis )0 (xit 1 0 12 1 yis + (xit xis )0 Figure 6: The Functions r (y1 ; y2 ; ) and R (y1 ; y2 ; ). Figure 7: The Functions E [r (y1 ; y2 ; )] and E [R (y1 ; y2 ; )]. 13 or with = (xit xis )0 , 0 = E [ (yit yis ) 1 f0 yit 1g 1 f0 = E [ (yit yis ) 1 f0 yit = E [ (yit yis ) 1f yit g 1 fyit = E [ (yit yis ) 1 fmax fyit yis + g 1 fyit 1g (xit 1g 1 f0 1 yis + g 1 fyis + g 1 f yis 1; yis g xis )] g 1f min fyit ; 1 1 yis gg (xit 1g (xit yis g (xit xis )] xis )] xis )] This is minus the derivative of E R yit ; yis ; (xit xis )0 b evaluated at b = , where 8 > (max fy1 1; y2 g) for > d < max fy1 1; y2 g > < R (y1 ; y2 ; d) = (y1 y2 d) for max fy1 1; y2 g d min fy1 ; 1 > > > : (min fy1 ; 1 y2 g) for d > min fy1 ; 1 y2 g y2 g As was the case for the censored model, the argument above only establishes that the true will solve the …rst order condition for minimizing E R yit ; yis ; (xit hand, it is clear that without additional strong assumptions, xis )0 b . On the other will not be the unique minimizer. The reason is that we know that, in general, the truncated regression model will not be identi…ed with exponentially distributed errors. As a result, assumptions must be added that rule out the exponential distribution. In Appendix 1 (section 7.2), we show that E R yit ; yis ; (xit xis )0 b , and more generally of E R yit ; yis ; (xit is the unique minimizer of xis )0 b , provided that the errors have a log– concave probability distribution. 3 Estimation The arguments leading to identi…cation of the parameters of interest, comparing two observations for the same individual and we showed that the unique minimizer of an expectation of the form E Q yit ; yis ; (xit Q. This suggests estimating , above were based on could be expressed as xis )0 b for some function by minimizing a sample analog of this such as n X Ti b = arg min 1 b n 2 1 i=1 X Q yit ; yis ; (xit xis )0 b 1 s<t Ti In this aggregation, observations get di¤erent weight depending on the number of observations for a given individual. Alternatively, one could also use objective functions of the type n arg min b 1X n X wist Q yit ; yis ; (xit i=1 1 s<t Ti 14 xis )0 b where the wist ’ are exogenous weights. In particular, with unbalanced panels, one might want wist s to depend on Ti , the number of time periods for individual i. For example, one can think of the usual …xed e¤ects estimator in a linear regression model as minimizing n X X i=1 1 s<t Ti 1 yis Ti so a simple natural choice for wist could be yit (xis xit )0 b 2 ; 1 Ti . For two– sided censoring, the resulting estimator is4 b = arg min b or more generally b = arg min b n X X wist U yit ; yis ; (xit xis )0 b X wist U xis )0 b i=1 1 s<t Ti n X yit ; yis ; (xit (15) i=1 1 s<t Ti where the functions U and U are de…ned in Section 2.1. Standard arguments5 yield Theorem 1 Consider a random sample of size n from Ti; fyit ; xit gTi : If t=1 1. yit = 2. ("i1 ; "i2 ; 8 > > > < > > > : 0 if x0 + "it if 0 it 1 if x0 + "it < 0 it x0 + "it it 1 ; x0 + "it > 1 it ; "iTi ) is continuously distributed conditional on Ti; fxit gTi t=1 with a density that is continuous and positive everywhere, 3. the sequence "i1 ; "i2 ; ; "iTi is stationary conditional on Ti; fxit gTi , and for any s; t t=1 there exists a random variable, st , i such that "is and "it are independent conditional on Ti st , i 4. the matrix E (xis xit ) (xis xit )0 1 < (xis xit )0 <1 has full rank 4 A Stata-program for calculating this estimator can be found at www.princeton.edu/~honore/stata. 5 Consistency follows from Theorem 4.1.1 of Amemiya (1985) and asymptotoc normality from, for example, The- orem 3.3 of Pakes and Pollard (1989). 15 then p where b is de…ned in (15) and = dE d n b 1 ! N 0; P s<t wi;t s u yit ; yis ; (xit db0 1 V xis )0 b (xit xis ) b= and 0 V = E vi vi with vi = X wi;t s xis )0 u yit ; yis ; (xit (xit xis ) : s<t These assumptions are consistent with a “…xed e¤ects” model in which "it = unrestricted and the sequence feit gTi " t=1 i + eit with " i independent and identically distributed. The assumptions also allow for some correlation in the eit ’ For example if (eis ; eit ) is bivariate normal with the " s. " " same variance, then they can be written as eis = Zis + Qi and eit = Zit + Qi where Zit , Zis and " " Qi are independent normals. So eis and eit are independent conditional on Qi . The assumption " " that ("i1 ; "i2 ; ; "iTi ) is continuously distributed is necessary if one wants to allow di¤erentiable(i.e., (d) = jdj). Without it, the derivative in the expression for to be non– might not exist. (d) = d2 , condition 3 can be reduced to assuming that the sequence "i1 ; "i2 ; When stationary conditional on Ti; fxit gTi t=1 ; "iTi is , and condition 2 is not necessary. In that case the terms in the asymptotic variance reduce to " X =E wi;t s 1 1 < (xis xit )0 <1 s<t 1 1 < (xis 1 yit < (xis xit )0 < yis 0 xit ) 1 1 0 < (xis <0 +1 1 yit < (xis xit )0 < yis 0 xit ) <1 (xis xit ) and 0 V = E vi vi with vi = X wi;t su yis ; (xis s<t 16 xit )0 (xis xit ) (xis 0 xit ) # Following standard arguments, these are consistently estimated by " n n o 1X X b= wi;t s 1 1 < (xis xit )0 b < 1 n i=1 s<t n o n o 1 1 < (xis xit )0 b < yis 1 1 0 < (xis xit )0 b < yis 1 n o n xit ) b < 0 + 1 1 0 yit < (xis and yit < (xis xit ) b < 1 0 o (xis xit ) (xis 0 xit ) # n 1X 0 b vi vi bb V = n i=1 with vi = b X wi;t s u yis ; (xis s<t xit )0 b (xis xit ) For two– sided truncation the resulting estimator is b = arg min b or more generally b = arg min b n X X wist R yit ; yis ; (xit xis )0 b (16) X wist R xis )0 b (17) i=1 1 s<t Ti n X yit ; yis ; (xit i=1 1 s<t Ti where the functions R and R are de…ned in Section 2.2. We have Theorem 2 Consider a random sample of size n from Ti; fyit ; xit gTi : If t=1 1. yit is drawn from the distribution of x0 + "it conditional on 0 it 2. ("i1 ; "i2 ; x0 + "it it 1 ; "iTi ) is continuously distributed conditional on Ti; fxit gTi t=1 with a density that is continuous and positive everywhere ; "iTi is stationary conditional on Ti; fxit gTi , and for any s; t t=1 3. the sequence "i1 ; "i2 ; there exists a random variable, density conditional on st , i such that "is and "it are independent and have log– concave st , i 4. the matrix E (xis Ti xit ) (xis xit )0 has full rank 17 1 < (xis xit )0 <1 then p where b is de…ned in (17) and = P dE d n b s<t wi;t s 1 ! N 0; r yit ; yis ; (xit db0 1 V xis )0 b (xit xis ) b= and 0 V = E vi vi with vi = X wi;t s r yit ; yis ; (xit xis )0 (xit xis ) s<t 4 Extensions 4.1 Mixed Censored/Truncation Having considered models with two– sided censoring or truncation, it is natural to also consider a regression model with censoring from one side and truncation from the other: yit = x0 + "it it (yit ; xit ) = (min fyit ; Uit g ; xit ) conditional on Lit yit (18) To simplify the notation, we again focus on the case where Lit = 0 and Uit = 1. In this case the moment condition based on the same logic as above is 0 = E 1 yit min yit x0 > it x0 is x0 ; 1 it = E 1 yit > (xit xis )0 min yit ; 1 + (xit = E t yit ; yis (xit x0 is xis )0 xis )0 1 yis x0 > is min yis 1 yis > x0 it x0 ; 1 is (xit min yis ; 1 x0 it xit ; xis xis )0 (xit xis )0 (xit xis )0 xit ; xis xit ; xis : where we have assumed that "it and "is are independent and identically distributed conditional on (xit ; xis ). This implies the unconditional moment condition E t yit ; yis (xit xis )0 18 (xit xis ) = 0 where t (y1 ; y2 ; d) = where e1 = c y2 , c2 = max fy1 8 > > > > > > > > > < > > > > > > > > > : 0 1 (y1 d < e1 c for for e1 < d < c2 c y2 y2 y1 d) for c2 < d < c3 for c3 < d < e4 c 1 0 e4 < d c for y2 ; y1 g and e4 = y1 . Note that c2 and c3 c 1; y2 g, c3 = min f1 are de…ned as before, but e1 and e4 di¤er from c1 and c4 . c c Let T (y1 ; y2 ; d) = 8 > > > > > > > > > < y2 ) e1 + 2 (1 c 2 (1 2 (1 y2 ) d + 2 (1 (y1 > > > > > > > > > : y2 ) c2 + (y1 y2 c2 )2 for y2 ) c2 + (y1 y2 c2 )2 d)2 y2 for e1 < d < c2 c for c2 < d < c3 2 (y1 1) d + 2 (y1 1) c3 + (y1 y2 c3 )2 2 (y1 1) e4 + 2 (y1 c 1) c3 + (y1 y2 c3 )2 for We then de…ne the estimator of for c3 < d < e4 c by minimizing XX i d < e1 c wits T yit ; yis ; (xit e4 < d c xis )0 b t<s The function T and its derivative are depicted in Figures 8 and 9 for a pair of uncensored observations and for a pair with one censored and one uncensored observation, respectively. Figure 10 shows the moment condition and the expected value of the objective function when y1 N (0:5; 1) and y2 N (0:4; 1) and both are truncated from below at 0 and censored from above at 1. As before, we also have E where t yit ; yis (xit is convex and symmetric, and ( ) = Let T (y1 ; y2 ; d) = 8 > > > > > > > > > < > > > > > > > > > : (1 (1 y2 ) e1 + (1 c y2 ) d + (1 (y1 xis )0 0( (xit xis ) = 0 ) (when it exists) y2 ) c2 + (y1 y2 c2 ) for y2 ) c2 + (y1 y2 c2 ) y2 d) for e1 < d < c2 c for c2 < d < c3 for c3 < d < e4 c (y1 1) d + (y1 1) c3 + (y1 y2 c3 ) (y1 1) e4 + (y1 c 1) c3 + (y1 y2 c3 ) for 19 d < e1 c e4 < d c Figure 8: The Functions t (y1 ; y2 ; ) and T (y1 ; y2 ; ). Neither Observation Censored. Figure 9: The Functions t (y1 ; y2 ; ) and T (y1 ; y2 ; ). One observation censored. 20 Figure 10: The Functions E [t (y1 ; y2 ; )] and E [T (y1 ; y2 ; )]. We then de…ne the estimator of by minimizing XX i 4.2 wits T yit ; yis ; (xit xis )0 b : t<s Pairwise Di¤erence Versions If we can estimate a panel data with 2 observations per unit, then we can apply the same idea to any two observations in a cross section, treating the constant in the cross sectional model as an individual– speci…c e¤ect. This idea was explicitly used in Honoré and Powell (1994) to construct estimators for the parameters of cross sectional (one– sided) censored and truncated regression models based on the panel data estimator in Honoré (1992). Among others, this also characterizes the relationship between the estimators in Manski (1987) and Han (1987) and between the estimators in Kyriazidou (1997) and Powell (1987). The same idea can be applied to the models with two– sided censoring and truncation considered here, and the asymptotic properties follow from the arguments used in Honoré and Powell (1994): 21 4.3 General Censoring Points. It is easy to generalize the results above to the case where the truncation and censoring points are not all 0 and 1. For example, consider the model 8 > > Li > < yit = x0 + "it > it > > : Ui then 8 > > > < yit Ui and we could estimate by Li = > Li > > : b = arg min b XX i with two– sided censoring if yit < Li if Li if xit Ui Li if "it Li Ui Li + 1 t<s yit < Li if Li if yit Ui wits U Ui yit > Ui 0 0 yit Li yis ; Li Ui yit Ui yit > Ui Li ; Li xit Ui xis Li 0 b (19) This simple approach does not work when the censoring points are time– varying, because then "it Lit Uit Lit is not stationary. In order to proceed, we need explicit expressions for the di¤erence in these “re– censored”residuals. We …rst note that only pairs for which the support of the re– censored residuals overlap can play a role in the moment conditions leading to the objective function. These pairs are characterized by Lit Uis < x0 it x0 < Uit is Lis and for such pairs, the di¤erence in the arti…cially censored residuals for individual i in periods t and s is mami Lis x0 ; yit is = mami Lis ; yit x0 it x0 ; Uis it x0 is x0 is mami Lit x0 ; yis it mami Lit ; yis + x0 it ; Uis x0 ; Uit is x0 is x0 it ; Uit + x0 it x0 is where we use the notation mami fa; x; bg = max fa; min fx; bgg. If we de…ne k (L; U; y; d) = and u (y1 ; y2 ; d; L1 ; L2 ; U1 ; U2 ) = 1 fL1 8 > > > < > > > : U y for d L d<y y for U2 < d < U1 22 U U <d<y d>y L L L2 g (k (L2 ; U2 ; y1 ; d) k (L1 ; U1 ; y2 ; d) + d) then E u yit ; yis ; x0 it x0 is ; Lit ; Lis ; Uit ; Uis (xit ; xis ) = 0 and hence E u yit ; yis ; x0 it x0 is ; Lit ; Lis ; Uit ; Uis (xit xis ) = 0: These will be the moment conditions that lead to the estimator in this case. Also de…ne K (L; U; y; d) = 8 > 2yU > > < > > > : 2yL U 2 for 2dU d)2 (y y L2 2dL d<y for U U <d<y d>y L L S (y1 ; y2 ; d; L1 ; L2 ; U1 ; U2 ) = K (L2 ; U2 ; y1 ; d) + K (L1 ; U1 ; y2 ; d) d2 and V (y1 ; y2 ; d; L1 ; L2 ; U1 ; U2 ) = and the estimator for 8 > S (y1 ; y2 ; L1 > > < U2 ; L1 ; L2 ; U1 ; U2 ) for S (y1 ; y2 ; d; L1 ; L2 ; U1 ; U2 ) for L1 > > > : S (y ; y ; U L2 ; L1 ; L2 ; U1 ; U2 ) for 1 2 1 d < L1 U2 U2 < d < U1 d > U1 L2 L2 is then de…ned by arg min b n XX wits V yit ; yis ; (xit xis )0 b; Lit ; Lis ; Uit ; Uis i=1 t<s A version of this can be developed for a general loss function. All of these extensions assume that the censoring and truncation points are exogenous in the sense that one must make assumptions on the error terms conditional on them. In a recent paper, Khan, Ponomareva, and Tamer (2011) consider a (one– sided) censored regression model with endogenous censoring. Their approach only leads to partial identi…cation, but it would be interesting to generalize it to more general versions of the models considered here. 5 Empirical Application In this section we apply the estimator in Section 2.1 to analyze the portfolio-reshu- ing e¤ect of a tax reform that increased the after-tax capital income on bonds relative to stocks in Denmark in 1987. We use a panel data set constructed from administrative records covering two years before and after the reform to estimate a portfolio share equation for bonds as a function of marginal tax 23 rates on capital income. The analysis presented here follows the literature on taxation and portfolio structure, e.g., Feldstein (1976), Hubbard (1985), King and Leape (1998), Samwick (2000), Poterba and Samwick (2002), Poterba (2002) and Alan, Crossley, Atalay, and Jeon (2010). These papers analyze (repeated) cross sections of households.6 Here the analysis is extended by using panel data and controlling for time– invariant correlated heterogeneity, i.e., …xed e¤ects. Controlling for correlated unobserved …xed factors is likely to be important in this context, since the portfolio composition of a household is likely to be in‡ uenced by time– invariant factors such as risk aversion and time discounting. In the next subsection we give a brief overview over the tax reform. After this, we introduce the data and present the results. 5.1 The Tax Reform The tax reform, announced in 1985 and implemented in 1987, broke the link between the marginal tax rates on earned income and capital income. Before the reform, all income was taxed at the same marginal tax rate. With the reform the tax rate on positive capital income for high-income households was decreased from 73 percent to 56 percent. The reform thereby increased the aftertax return on interest-bearing assets and therefore encouraged households to shift their portfolios toward such assets. The reform also changed the tax value of interest deductions from 73 to about 50 percent, and this substantially increased the cost of debt, primarily mortgages, for leveraged high-income households. For such households the reform e¤ectively brought a negative wealth shock, giving them a strong incentive to lower their debt burden.7 The exact changes, however, di¤ered across municipalities. The Danish income tax system is built around a proportional local government tax and a progressive tax collected by the central government. While the progressive schedule is the same for everybody in Denmark, the local 6 Bakija (2000) uses the limited panel module of the American Survey of Consumer Finances (SCF) to study portfolio changes around the 1988 tax reform. However, his data set is very small (984 households) and unrepresentative due to the well-known attrition problem in the SCF panel module; see Kennickell and Woodburn (1997). More important in this context, the estimators applied do not exploit the full potential of the panel data in handling unobserved heterogeneity. Ioannides (1992) also employs the 1983-1986 SCF panel module but does not control for unobserved heterogeneity. 7 Alan and Leth-Petersen (2006) document that the reduced value of the interest deduction led households to liquidate …nancial assets to lower their mortgage debt. This was possible because pre-payment of mortgage debt is not restricted in Denmark. 24 Figure 11: Marginal Tax Rate for High– Tax Municipality. government tax rates vary across municipalities. A tax ceiling, however, insured that the marginal tax rate could be at the maximum 73 percent. After the reform the tax ceiling on earned income was reduced to 68 percent in the highest bracket8 and 56 percent in the middle bracket. Capital income was now taxed at the same rate independently of the level of earned income. The marginal tax rates across tax brackets before and after the reform are summarized in Table 1 (see Appendix 2) The application of a tax ceiling together with the heterogeneous local government tax rates implies that the reform had di¤erential e¤ects on people living in di¤erent municipalities. Figures 11 and 12 illustrate the changes in marginal tax rates due to the reform for a high-tax and a low-tax municipality, respectively. For a high-income person living in the municipality with the high local government tax, the marginal tax rate on positive net capital income falls by 14.5 percentage points and the marginal tax rate on negative net capital income falls by 20.5 percentage points. For a similar person living in the municipality with the low local government tax rate, the marginal tax rate on positive 8 Approximately 20 percent of the population belong in the top bracket. 25 Figure 12: Marginal Tax Rate for Low– Tax Municipality. net capital income falls by 16.1 percentage points and negative net capital income falls by 22.1 percentage points. It is these di¤erences in changes of marginal tax rates that we will exploit for identifying the e¤ect of changes in marginal tax rates on the portfolio allocation when using the …xed e¤ects estimator. The marginal tax rates on capital income refer to income received in the form of dividends on stocks and interest payments from interest bearing accounts and bonds. Both before and after the reform, realized capital gains/losses associated with trading assets were generally not taxed. The exemption from this rule is capital gains from corporate stocks held for less than three years. Such capital gains are taxed as earnings. Dividend payments were low relative to interest received from bonds.9 This suggests that lowering the marginal tax rate on positive capital income a¤ected bonds and stocks di¤erentially, favoring mainly income from bonds. In the empirical analysis we therefore focus on reshu- ing between bonds and stocks. 9 The median household in the sample holding stocks received dividends corresponding to 2 percent of the value of the stocks. The median household in the sample holding bonds received interest payments from these corresponding to 10 percent of the value of the bonds. 26 5.2 Data The data set is drawn from a random sample of 10 percent of the Danish population observed in the years 1984 to 1988. Information on portfolio allocations, income, wealth and demographics is collected and merged from di¤erent public administrative registers for all adult members of the household that the sampled person belongs to. Portfolio and income information is obtained from the income tax register. The portfolio information exists because Denmark had a wealth tax that required all wealth holdings to be reported to the tax authorities. This information allows us to break the wealth of each household into holdings of stocks and bonds. “Stocks” includes all holdings of publicly and privately traded stocks, and “bonds” includes government and corporate bonds. The holdings of stocks and bonds are self– reported through the tax return and then audited by the tax authority. 5.2.1 Sample selection For our analysis we exclude observations if one of the household members is self-employed, since register data are not likely to contain a good measure of own business wealth and because taxable income is quite volatile for those individuals. Sampled individuals younger than 18 or older than 60 are dropped as are students and individuals living together with his/her parents or living in a common household, i.e., a household with more than one family. To keep the focus on the importance of tax incentives, we include only stable couples, i.e., couples where the partner is the same in 1984 through 1988. On the same grounds we also exclude couples moving in the sample period. For the purpose of the analysis we require that households entering the sample be observed in all years in the period 1984-1988 so that we have a balanced panel. Our objective is to investigate whether households reshu- e their portfolios in response to a change in tax incentives. As in most industrialized countries many Danish households have fairly undiversi…ed portfolios. Since the decrease in the value of interest deductions generated a large negative wealth shock, clearly, these households are not likely to engage in portfolio reshu- ing and hence cannot give us a clean answer regarding portfolio readjustments. We therefore construct a sub-sample of households holding positive amounts of stocks or bonds of at least 5,000 DKK in 1984. We also require households to hold a positive amount of either stocks or bonds throughout the rest of the observation period. This selection is introduced because we want to focus on households with a potential to reshu- e between stocks and bonds. Also renters are deselected because there 27 are few renters with diversi…ed portfolios.10 The …nal subsample includes 8,577 households.11 5.3 Results In this section we investigate if households reshu- ed their portfolio of bonds and stocks as a response to the changes in relative after-tax returns on assets brought about by the 1987 tax reform. To do, this we employ the estimator presented above and estimate a portfolio share equation where the fraction of bonds in …nancial wealth, de…ned as the sum of bonds and stocks, is regressed on the marginal tax rate on positive capital income and some control variables. The distinguishing feature of our data set is the panel dimension. This facilitates estimating portfolio share equations allowing for correlated unobserved time– invariant heterogeneity. This is important because we believe that unobserved time– invariant factors, such as risk aversion and time preferences, are correlated with wealth. High risk aversion may, for example, lead to a higher portfolio share of safe assets, such as bonds, for a given level of wealth. Before the reform, capital income and earnings were lumped together and taxed according to a progressive tax scheme. This implies that households choose their tax bracket when choosing their portfolios and that the marginal tax rate on capital income is likely to be an endogenous regressor. We address this by calculating the marginal tax rate on capital income based on the household’ s income in 1984, the year before the reform was announced, but using current year rules. In this way the individual level tax bracket is allocated based on information that was predetermined relative to the portfolio response to the reform. We regress portfolio shares on the marginal tax rate on positive capital income, the log of total …nancial assets, i.e., assets held in stocks and bonds, and a set of year dummies. Tax rate changes vary across municipalities, but most of the change in tax rates is common across municipalities. Year dummies control for the e¤ect of this common part, thereby also removing the major part of the wealth e¤ect brought about by the reform. E¤ectively, by introducing year dummies, the coe¢ cients on marginal tax rates are identi…ed by di¤erences in changes of marginal tax rates. Year dummies may also pick up common e¤ects related to ‡ uctuations in assets. Financial assets 10 For assessing portfolio reshu- ing renters could have been included. We have chosen to leave them out of this analysis because there are only a few renters (898) with positive …nancial wealth of at least 5000 DKK in 1984. Moreover, renters generally do not provide a good comparison group for homeowners, since di¤erent preference parameters may govern their behavior. 11 See Alan and Leth-Petersen (2006) for a more detailed analysis. 28 control for any remaining wealth e¤ect that might be present.12 Table 2 of Appendix 2 presents the parameter estimates from estimating random e¤ects Tobit and …xed e¤ects censored regression models for the portfolio share of bonds in …nancial wealth. The estimated parameter on the marginal tax rate is negative.13 If year dummies and …nancial assets pick up the wealth e¤ect related to the reform, in particular the e¤ect of the reduction in the value of interest deduction that led households to liquidate …nancial assets, then this is exactly what economic theory predicts. Households should substitute from stocks toward bonds, whose relative after-tax return increased, and this is what the results indicate. Considering the corresponding random e¤ects estimates, we can see that the parameter estimates on …nancial assets and on year dummies are quite di¤erent, and the test of equality of all the parameters in the random e¤ects and …xed e¤ects speci…cations rejects. 6 Concluding Remarks This paper constructs estimators for panel data regression models with individual speci…c heterogeneity and two– sided censoring and truncation. Following Powell (1986) the estimation strategy is based on moment conditions constructed from re– censored or re– truncated residuals. While these moment conditions do not identify the parameter of interest, they can be used to motivate objective functions that do. This part is the main methodological contribution of the paper. We apply one of the estimators to study the e¤ect of a Danish tax reform on household portfolio choice. We …nd that a random e¤ects speci…cation can be rejected in favor of the “…xed”e¤ects speci…cation studied here, although both models yield the same sign of the key parameter that one would anticipate from economic theory. The estimators are fairly easy to implement and a link to a program that calculates the leading estimator is provided at http://www.princeton.edu/~honore/stata. 12 An alternative identi…cation strategy could be based on comparing the behavior of households in di¤erent tax brackets. Households in the lowest tax bracket faced only very small changes in marginal tax rates on capital income, and households in the middle tax bracket faced di¤erent changes in marginal tax rates than households in the highest tax bracket. In our case this is not a natural approach to follow. High–and low– income people are di¤erent in terms of wealth levels and portfolio composition and possibly di¤erent with respect to preference parameters such as the discount rate and the level of risk aversion. Households in lower tax brackets therefore do not represent a natural control group for high– income households. 13 As explained in Honoré (2008), the parameter estimates for both the random e¤ects and the …xed e¤ects models can be converted to marginal e¤ects by multiplying them by the fraction of observations that are not censored. 29 References Abrevaya, J. (1999): “Rank estimation of a transformation model with observed truncation,” Econometrics Journal, 2(2), 292– 305. Alan, S., T. Crossley, K. Atalay, and S.-H. Jeon (2010): “New Evidence on Taxes and Portfolio Choice,” Journal of Public Economics, 94, 813– 823. Alan, S., and S. Leth-Petersen (2006): “Tax Incentives and Household Portfolios: A Panel Data Analysis,” Center for Applied Microeconometrics, University of Copenhagen, Working paper number 2006-13. Amemiya, T. (1985): Advanced Econometrics. Harvard University Press. Andrews, M., T. Schank, and R. Simmons (2005): “Does Worksharing Work? Some Empirical Evidence from the IAB-Establishment Panel,” Scottish Journal of Political Economy, 52(2). Arellano, M., and B. E. Honoré (2001): “Panel Data Models. Some Recent Developments,” Handbook of Econometrics, 5, Elsevier Science, Amsterdam. Bakija, J. (2000): “The E¤ect of Taxes on Portfolio Choice: Evidence from Panel Data Spanning the Tax Reform Act of 1986,” Williams College. Charness, G., G. Frechette, and J. Kagel (2004): “How Robust Is Laboratory Gift Exchange?,” Experimental Economics, 7(2), 189– 205. Chen, S. (forthcoming): “Nonparametric Identi…cation and Estimation of Truncated Regression Models,” Review of Economic Studies. de Figueriredo, J., and E. Tiller (2001): “The Structure and Conduct of Corporate Lobbying: How Firms Lobby the Federal Communications Commission,” Journal of Economics and Management Strategy, 10(1), 91– 122. Fehr, E., E. Kirchler, A. Weichbold, and S. Gachter (1998): “When Social Norms Overpower Competition: Gift Exchange in Experimental Labor Markets,” Journal of Labor Economics, 16(2), 324– 351. Feldstein, M. (1976): “Personal Taxation and Portfolio Composition: An Econometric Analysis,” Econometrica, 44(4), 631– 650. 30 Fenn, G., and N. Liang (2001): “Corporate Payout Policy and Managerial Stock Incentives,” Journal of Financial Economics, 60(1), 45– 72. Ferrie, J. P., and K. Rolf (2011): “Socioeconomic Status in Childhood and Health after Age 70: A New Longitudinal Analysis for the U.S., 1895-2005,” SSRN eLibrary. Gifford, K., and J. Bernard (2005): “In‡ uencing Consumer Purchase Likelihood of Organic Food,” International Journal of Consumer Studies, pp. 1– 9. Han, A. (1987): “Nonparametric Analysis of a Generalized Regression Model: The Maximum Rank Correlation Estimator,” Journal of Econometrics, 15. Honoré, B. E. (1992): “Trimmed LAD and Least Squares Estimation of Truncated and Censored Regression Models with Fixed E¤ects,” Econometrica, 60, 533– 565. (2008): “On Marginal E¤ects in Semiparametric Censored Regression Models,”Princeton University. Honoré, B. E., and L. Hu (2004): “Estimation of Cross Sectional and Panel Data Censored Regression Models with Endogeneity,” Journal of Econometrics, 122(2), 293– 316. Honoré, B. E., and J. L. Powell (1994): “Pairwise Di¤erence Estimators of Censored and Truncated Regression Models,” Journal of Econometrics, 64, 241– 78. Honoré, B. E., and J. L. Powell (1994): “Pairwise Di¤erence Estimators of Censored and Truncated Regression Models,” Journal of Econometrics, 64(1-2), 241– 78. Houston, J., and M. Ryngaert (1997): “Equity Issuance and Adverse Selection: A Direct Test Using Conditional Stock O¤ers,” Journal of Finance, 52(1), 197– 219. Hu, L. (2002): “Estimation of a Censored Dynamic Panel Data Model,” Econometrica, 70(6), pp. 2499– 2517. Huang, M., and R. Hauser (1998): “Trends in Black-White Test Score Di¤erences: WORDSUM Vocabulary Test,” in The Rising Curve: Long-IQ and Related Measures, ed. by U. Neisser, pp. 303– 332. American Psychological Association, Washington, D.C. (2001): “Convergent Trends in Black-White Verbal Test Score Di¤erentials in the U.S.: Period and Cohort Perspectives,” EurAmerica, 31(2), 185– 230. 31 Hubbard, R. G. (1985): “Personal Taxation, Pension Wealth, and Portfolio Composition,” The Review of Economics and Statistics, 67(1), 53– 60. Ioannides, Y. (1992): “Dynamics of the Composition of Household Asset Portfolios and Life Cycle,” Applied Financial Economics, 2(3), 145– 159. Kennickell, A., and L. Woodburn (1997): “Weighting Design for the 1983-89 SCF Panel,” Washington DC: Federal Reserve Board of Governors. Khan, S., M. Ponomareva, and E. Tamer (2011): “Identi…cation of Panel Data Models with Endogenous Censoring,” MPRA Paper 30373, University Library of Munich, Germany. King, M., and J. Leape (1998): “Wealth and Portfolio Composition: Theory and Evidence,” Journal of Public Economics, 69(2), 155– 193. Kyriazidou, E. (1997): “Estimation of a Panel Data Sample Selection Model,”Econometrica, 65, 1335– 1364. Lafontaine, F. (1993): “Contractual Arrangements as Signalling Devices: Evidence from Franchising,” Journal of Law, Economics, and Organization, 9(2), 256– 289. Manski, C. (1987): “Semiparametric Analysis of Random E¤ects Linear Models from Binary Panel Data,” Econometrica, 55, 357– 62. McMillan, J., and C. Woodruff (1999): “Inter…rm Relationships and Informal Credit in Vietnam,” Quarterly Journal of Economics, 114(4), 1285– 1320. Nickerson, J., and B. Silverman (2003): “Why Aren’ All Truck Drivers Owner-Operators? t Asset Ownership and the Employment Relation in Interstate For-Hire Trucking,” Journal of Economics and Management Strategy, 12(1), 91– 118. Officer, M. (2004): “Collars and Renegotiation in Mergers and Acquisitions,”Journal of Finance, 59(6), 2719– 2743. Pakes, A., and D. Pollard (1989): “Simulation and the Asymptotics of Optimization Estimators,” Econometrica, 57, 1027– 1057. Petersen, M., and R. Rajan (1994): “The Bene…ts of Lending Relationships: Evidence from Small Business Data,” Journal of Finance, 49(1), 3– 37. 32 (1995): “The E¤ect of Credit Market Competition on Lending Relationships,” Quarterly Journal of Economics, 110(2), 407– 443. Poterba, J., and A. Samwick (2002): “Taxation and Household Portfolio Composition: US Evidence from the 1980s and 1990s,” Journal of Public Economics, 87(1), 5– 38. Poterba, J. M. (2002): “Taxation and Portfolio Structure: Issues and Implications,”in Household Portfolios, ed. by L. Guiso, M. Haliassos, and T. Jappelli. MIT Press. Powell, J. L. (1986): “Symmetrically Trimmed Least Squares Estimation for Tobit Models,” Econometrica, 54(6), 1435– 60. (1987): “Semiparametric Estimation of Bivariate Latent Models,” Working Paper no. 8704, Social Systems Research Institute, University of Wisconsin– Madison. Samwick, A. (2000): “Portfolio Responses to Taxation: Evidence from the End of the Rainbow,” in Does Atlas Shrug? The Economic Consequences of Taxing the Rich, ed. by J. Slemrod. Harvard University Press, Cambridge, MA. 33 7 Appendix 1: Proofs and Derivations 7.1 Two– Sided Censoring This section provides justi…cation for the statements about the estimators for the model with two– sided censoring. We …rst verify that u (yit ; yis ; x0 it x0 ) in equation (8) does indeed yield the is di¤erence in the re-censored residuals de…ned in (6) and (7). Write d = x0 it x0 , and consider is …rst the case where d > 0. In this case, the di¤erence in the re-censored residuals is max fyit d; 0g min fyis ; 1 dg which is most easily analyzed by considering a number of cases for d between 0 and 1. As mentioned earlier, the contribution to the moment condition should be 0 when d is outside the interval between 1 and 1,which is consistent with the de…nition of u: There are four cases based on combinations of whether yit Case 1 (yit (yit d) d (1 Case 2 (yit 0 (1 d) = yit d d) = d Case 3 (yit (yit d) 0 and yis 1 d:14 1 d): In this case, max fyit d; 0g min fyis ; 1 dg = 1 d): In this case, max fyit d; 0g min fyis ; 1 dg = 1 d): In this case, max fyit d; 0g min fyis ; 1 dg = 1. d d 0 and yis 1. 0 and yis 0 and yis yis = yit Case 4 (yit d yis 0 and yis d. 1 d): In this case, max fyit d; 0g min fyis ; 1 dg = 0 yis = yis . Case 3 corresponds to values of d close to (or at) 0. Speci…cally, the region for Case 3 is (0; min fyit ; 1 yit yis yis g) = (0; c3 ). Noting that c2 0, it is clear that max fyit d; 0g min fyis ; 1 d is consistent with the de…nition of u in equation (8). Case 2 corresponds to values of d close to (or at ) 1, speci…cally the region (max fyit ; 1 (c4 ; 1), and it is again clear that u delivers max fyit d; 0g The region that de…nes the other two cases is (min fyit ; 1 min fyis ; 1 yis g ; 1) = dg. yis g ; max fyit ; 1 Cases 1 and 4 give di¤erent expressions depending on whether yit or 1 14 dg = yis g) = (c3 ; c4 ). yis is larger, but these Since both the di¤erence in the re-censored residuals and u are continuous in d, it is not necessary to distinguish between closed and open intervals in the following discussion. 34 expressions correspond exactly to the two cases for the max in the de…nition of u in (8) over the interval (c3 ; c4 ). The case d < 0 is dealt with in exactly the same manner. We now turn to the question of why xis )0 b E U yit ; yis ; (xit is uniquely minimized at the true . This is the key for consistency of the proposed estimator. The result follows from the following lemma Lemma 3 Suppose yi1 = mami f0; + "i1 ; 1g and yi2 = mami f0; "i2 ; 1g where "i1 and "i2 are identically distributed random variables with support on the whole real line. Then arg max E [U (yi1 ; yi2 ; d)] = d2[ 1;1] Proof: For 1 d 0 E [u (yi1 ; yi2 ; d)] = E [max fyi1 d; 0g 8 > > > < 1 if > > > : 1 min fyi2 ; 1 = E [max fmami f0; + "i1 ; 1g = E [max fmami f0 = E [mami f0; d; if 1< if 1 d; 0g dg <1 dg] d + "i1 ; 1 d + "i1 ; 1 1 min fmami f0; "i2 ; 1g ; 1 dg] dg ; 0g dg] mami f0; "i2 ; 1 mami f0; "i2 ; 1 dg] If "i1 (and "i2 ) have full support, then this is negative for d > and positive for d < . For 1 d 0 E [u (yi1 ; yi2 ; d)] = E [min fyi1 ; 1 + dg max fyi2 + d; 0g] = E [min fmami f0; + "i1 ; 1g ; 1 + dg max fmami f0; "i2 ; 1g + d; 0g] = E [mami f0; + "i1 ; 1 + dg max fmami fd; "i2 + d; 1 + dg ; 0g] = E [mami f0; + "i1 ; 1 + dg mami f0; "i2 + d; 1 + dg] 35 If "i1 (and "i2 ) have full support, then this is negative for d > and positive for d < . Since E U 0 (yi1 ; yi2 ; d) = E [u (yi1 ; yi2 ; d)] the argument above shows that arg max E [U (yi1 ; yi2 ; d)] = d2[ 1;1] Corollary 4 Consider the model 8 > > > < 1 if 1 if > > > : 1 1< if <1 1 yit = x0 + "it it yit 8 > 0 if > yit < 0 > < = y if 0 yit 1 > it > > : 1 if yit > 1 for t = 1; 2. If "it is stationary conditional on (xi1 ; xi2 ) with support on the whole real line, then the set of solutions to max E U yi1 ; yi2 ; mami b xi2 )0 b; 1 1; (xi1 is b : P mami 1; (xi1 xi2 )0 b; 1 = mami 1; (xi1 xi2 )0 ; 1 =1 The Corollary above requires that the errors are stationary conditional on the regressors. This is much more general than the usual assumption that the individual— speci…c e¤ect and the contemporaneous errors interact additively. To see that E U yit ; yis ; (xit xis )0 b in (12) is uniquely minimized, it is convenient to assume that "it and "is are independent and identically distributed conditional on some individual— speci…c e¤ect, vi . With this assumption, the argument for why E U yit ; yis ; (xit Speci…cally, when 1 d xis )0 b is uniquely minimized follows essentially the same logic as above. 0 (u (yit ; yis ; d)) = (mami f0; Let = (xit d + "it ; 1 xis )0 . 36 dg mami f0; "is ; 1 dg) If "it and "is are independent and identically distributed conditional on vi , then E [ (u (yit ; yis ; ))j xit ; xis ] = E [ E [ (u (yit ; yis ; ))j = E [ E [ (mami f0; "it ; 1 because mami f0; "it ; 1 g g mami f0; "is ; 1 i ; xit ; xis ]j xit ; xis ] mami f0; "is ; 1 g)j i ; xit ; xis ]j xit ; xis ] =0 g is symmetrically distributed conditional on i, and ( ) is an odd function. For d > mami f0; d + "it ; 1 dg mami f0; "is ; 1 dg mami f0; "it ; 1 with probability 1 (conditional on E [ (mami f0; d + "it ; 1 dg i ), and since mami f0; "is ; 1 dg is increasing mami f0; "is ; 1 E [ (mami f0; "it ; 1 The line of argument is the same when d < dg dg)j dg i ; xit ; xis ] mami f0; "is ; 1 dg)j i ; xit ; xis ] =0 . Strict inequalities follow from a full support assumption on "i1 (and "i2 ). We therefore have that E [ U (yit ; yis ; d)j xit ; xis ] is decreasing to the left of (xit increasing to the right. Hence it is minimized at d = (xit this implies that E U 7.2 yit ; yis ; (xit xis )0 and xis )0 . Subject to a rank condition, xis )0 b xit ; xis is minimized at b = : Two– Sided Truncation The following Lemma (combined with the obvious rank– condition) establishes that minimization of E R yi1 ; yi2 ; (xi1 x2i )0 b will identify if the distribution of " is log– concave. This is the assumption that was made in a number of other papers (including Honoré (1992), Honoré and Powell (1994) and Abrevaya (1999); see also the discussion in Chen (forthcoming)). In Section 2.2, we only consider the case with (two– sided) truncation at 0 and 1. It is just as easy to prove identi…cation for general individual– and time– speci…c truncation points. In the following we therefore denote the truncation points by L and U . Lemma 5 Let (L; U ) be a vector of random variables such that L < U with probability 1. Assume that " is independent of (L; U ) and has a continuous, log– concave distribution with support on the whole real line. Let yit = it +"it for some real number, 37 it , and consider two draws (yi1 ; Li1 ; Ui1 ) and (yi2 ; Li2 ; Ui2 ) from the distribution of (y; L; U ) conditional on L < y < U . then E [R (yi1 ; yi2 ; d)] is uniquely minimized at d = 7.3 i2 . i1 Proof of Lemma 5. Let E denote expectation conditional on truncation and E in population. The moment condition can then be written as E [ (4yi 4di ) 1 fLi2 = E [ (4yi di2 yi1 4di ) 1 fLi2 + 4di = E[ (4"i + 4 1 fLi1 + (4 where 4ai = ai1 4di ) i i ) 1 fLi2 = E [ (4"i + yi1 P (Li1 yi1 i 4di ) i i i "i1 + yi2 4di Ui2 yi2 di2 Ui1 (4 i + i Ui1 di1 g] 4di g] 4di )g 4di )g] i Ui2 i2 di1 4di ) "i1 i i g 1 fLi1 i2 Ui2 i2 i1 i g 1 fLi1 i1 + "i2 "i2 i Ui1 i1 Ui1 + i1 + Ui1 g 1 fLi2 yi2 Ui2 g] 1 : Ui1 ; Li2 yi2 Ui2 ) It su¢ ces to show that this is nonpositive for left of 0, 0 for di2 g 1 fLi1 Ui2 + 4di g 1 fLi1 "i1 i i2 Ui2 Ui1 + (4 i1 = (4 i2 i ) 1 fLi2 1 fLi1 (4 "i2 + ai2 . Letting = E[ (4"i + yi1 4di ) 1 fLi2 i di1 = 0; nonnegative for i i < 0, strictly negative in a neighborhood to the > 0, and strictly positive in a neighborhood to the right of 0. Now consider the term E [ (4"i + 1 fLi1 i1 i ) 1 fLi2 "i1 De…ning wi1 = 1 2 Ui1 ("i1 i2 i "i1 bj i1 g 1 fLi2 "i2 ) and wi2 = i2 1 2 i g 1 fLi1 i2 "i2 bj i1 + i "i2 Ui1 i1 38 ig i2 g]: ("i1 + "i2 ) (so "i1 = (wi1 + wi2 ) and "i2 = wi2 and 4"i = 2wi1 ) + wi1 i g] ig E [ (4"i + 1 fLi1 i ) 1 fLi2 "i1 i1 = E [ (2wi1 + i2 Ui1 Ui2 i1 g 1 fLi2 i ) 1 fLi2 + "i1 i i2 wi2 1 fLi1 i1 1 fLi1 i1 wi1 + wi2 Ui1 1 fLi2 i2 wi2 Ui2 = E [ (2wi1 + 1 Li2 "i2 i Ui1 i1 + ig i2 g] ig i2 1 Ui1 + 1 Li1 1 = E [ (2wi1 + 1 2 min Ui2 = E = E i2 1 2 + wi2 1 2 i wi1 + 1 2 1 2 i wi1 + i i + wi2 + 1 2 1 2 wi1 + i 1 2 wi1 + i Ui2 Li1 + i Ui1 i1 i2 i 1 2 + wi2 wi2 + i1 Li2 + i 1 2 wi2 i2 1 2 i 1 2 i i + wi2 + ] i) wi2 i2 1 i ; Ui1 + 2 i1 1 i ; Li1 2 + wi2 wi2 + i1 1 i ; Ui2 + 2 i2 1 2 i i g] + wi2 + 1 2 i + wi2 + i i2 (2wi1 + E 1 2 wi2 + i1 Ui2 + 1fmax Li2 wi1 + i1 ig i) wi2 i2 i1 + Ui2 + i2 g] wi1 Ui1 Ui2 i1 i1 g i wi1 "i2 i2 wi1 + wi2 i i g 1 fLi1 i2 wi2 i) 1 (2wi1 + where ci = min Ui2 i2 1 i ; Li1 + 2 1 ci wi1 + 2 i) 1 wi2 ci i1 1 i ; Ui1 2 i1 wi2 + 1 i ; Li2 + 2 i2 + wi2 + 1 2 ci i wi1 + 1 2 i; + wi2 1 2 Li1 + ci i i1 + wi2 wi2 ; Li1 ; Ui1 ; Li2 ; Ui2 1 2 i ; Ui1 i1 wi2 + 1 2 i; Li2 + i2 Strict log– concavity of "it , implies that wi1 is strictly unimodal and symmetric conditional on wi2 . 39 . It therefore follows that E E (2wi1 + i) 1 ci wi1 + 1 2 ci i 8 > 0 > > > > > > > <0 > < =0 > > > > > 0 > > > > : >0 wi2 ; Li1 ; Ui1 ; Li2 ; Ui2 if i <0 if i < 0, if i =0 if if P ci wi1 + 1 2 i ci > 0 i >0 i > 0, P ci wi1 + 1 2 i ci > 0: Since the "’ (and hence the wi ’ are continuous, the condition that P s s) ci wi1 + 1 2 i ci > 0 will be satis…ed if P (ci > 0) > 0. We will next show that for 1 fci > 0g = 1 1 = 1 1 = 1 i 2 (0; k), P (ci > 0) > 0 for some k > 0. Note that 1 1 Ui1 wi2 + i >0 i1 2 1 Li2 + i2 + wi2 + 1 Ui2 wi2 i >0 i2 2 1 1 wi2 > Li1 1 Ui1 i1 + i i1 + i > wi2 2 2 1 1 wi2 > Li2 1 Ui2 i2 i i2 i > wi2 2 2 1 1 Ui1 1 Ui2 i1 + i > wi2 > Li1 i1 + i 2 2 Li1 + i1 + wi2 1 2 i 1 2 i2 >0 i >0 1 2 i > wi2 > Li2 i2 This will have positive probability provided that P Li1 i1 + 1 2 i < Ui2 i2 1 2 >0 which follows from Lemma 6. Lemma 6 If (U1 ; V1 ) and (U2 ; V2 ) are two independent draws of a random vector (U; V ) with P (U < V ) = 1, then there exists a k > 0 such that for 0 < Proof. 1 2. Since V ) > 0. U > 0 with probability 1, there exists an m > 0 such that P (V Now consider the space f(u; v) : v U > m) u > mg. This can be divided into a countable number of regions, Ak , such that for (u1 ; v1 ) ; (u2 ; v2 ) 2 Ak , ju1 P (V < k, P (U1 < V2 u2 j < m 2 and jv1 v2 j < m 2. Since U > m) = 1 , at least one of these regions, Ak , must have positive probability. Hence there 2 is positive probability that (U1 ; V1 ) and (U2 ; V2 ) in the statement of the lemma both belong to this Ak , in which case U1 < V1 m < V2 m 2. 40 1 2 i : 8 Appendix 2: Empirical Results Table 1: Marginal tax rates before and after implementation of the 1987 tax reform Before Reform After Reform Tax bracket Earnings + Cap inc 0-113 M + 19:75 0 113-186 M + 34:15 130 186- M + 44:95 Tax ceiling 73 Earnings inc < 0 inc: > 0(1) 130 M + 22:00 M + 22:00 M + 22:00 200 M + 28:00 M + 22:00 M + 28:00 M + 40:00 M + 22:00 M + 28:00 Tax bracket 200 Tax ceiling 68:00=56:00(2) 56:00 Note: M is the local government tax rate. Threshold values for the tax brackets are given in 1000 DKK. Thresholds are adjusted yearly. Threshold values used in the table are for 1986 (before the reform) and 1987 (after the reform). The marginal tax rates refer to personal income (as opposed to household income). (1) The tax brackets for positive net capital income refer to the sum of earnings and positive net capital income. After the reform positive capital income is taxed progressively up to the …rst threshold, 130,000 DKK. For a married couple the progression threshold is 260,000 based on the sum of their joint positive net capital income and earnings. (2) The 68 percent tax ceiling applies only to the top bracket. 41 Table 2:. Random and Fixed E¤ects Censored Regression Estimates of the Portfolio Share of Bonds in Financial Wealth. Fixed E¤ects Random E¤ects 0:130 0:206 (0:071) (0:045) 0:177 0:085 (0:008) (0:003) 0:214 0:157 (0:006) (0:006) 0:314 0:248 (0:009) (0:006) 0:318 0:285 (0:013) (0:008) 0:383 0:331 (0:013) (0:008) MTR capital income Ln(Financial Assets) D85 D86 D87 D88 Constant — 0:047 (0:046) # households/observations 8,577 / 42,885 # left/right censored obs 7,529 / 15,655 Test of Parameter Equality (d.f.) 167 (6) Standard errors in parenthesis 42 Working Paper Series A series of research studies on regional economic issues relating to the Seventh Federal Reserve District, and on financial and economic topics. A Leverage-based Model of Speculative Bubbles Gadi Barlevy WP-08-01 Displacement, Asymmetric Information and Heterogeneous Human Capital Luojia Hu and Christopher Taber WP-08-02 BankCaR (Bank Capital-at-Risk): A credit risk model for US commercial bank charge-offs Jon Frye and Eduard Pelz WP-08-03 Bank Lending, Financing Constraints and SME Investment Santiago Carbó-Valverde, Francisco Rodríguez-Fernández, and Gregory F. Udell WP-08-04 Global Inflation Matteo Ciccarelli and Benoît Mojon WP-08-05 Scale and the Origins of Structural Change Francisco J. Buera and Joseph P. Kaboski WP-08-06 Inventories, Lumpy Trade, and Large Devaluations George Alessandria, Joseph P. Kaboski, and Virgiliu Midrigan WP-08-07 School Vouchers and Student Achievement: Recent Evidence, Remaining Questions Cecilia Elena Rouse and Lisa Barrow WP-08-08 Does It Pay to Read Your Junk Mail? Evidence of the Effect of Advertising on Home Equity Credit Choices Sumit Agarwal and Brent W. Ambrose WP-08-09 The Choice between Arm’s-Length and Relationship Debt: Evidence from eLoans Sumit Agarwal and Robert Hauswald WP-08-10 Consumer Choice and Merchant Acceptance of Payment Media Wilko Bolt and Sujit Chakravorti WP-08-11 Investment Shocks and Business Cycles Alejandro Justiniano, Giorgio E. Primiceri, and Andrea Tambalotti WP-08-12 New Vehicle Characteristics and the Cost of the Corporate Average Fuel Economy Standard Thomas Klier and Joshua Linn WP-08-13 Realized Volatility Torben G. Andersen and Luca Benzoni WP-08-14 Revenue Bubbles and Structural Deficits: What’s a state to do? Richard Mattoon and Leslie McGranahan WP-08-15 1 Working Paper Series (continued) The role of lenders in the home price boom Richard J. Rosen WP-08-16 Bank Crises and Investor Confidence Una Okonkwo Osili and Anna Paulson WP-08-17 Life Expectancy and Old Age Savings Mariacristina De Nardi, Eric French, and John Bailey Jones WP-08-18 Remittance Behavior among New U.S. Immigrants Katherine Meckel WP-08-19 Birth Cohort and the Black-White Achievement Gap: The Roles of Access and Health Soon After Birth Kenneth Y. Chay, Jonathan Guryan, and Bhashkar Mazumder WP-08-20 Public Investment and Budget Rules for State vs. Local Governments Marco Bassetto WP-08-21 Why Has Home Ownership Fallen Among the Young? Jonas D.M. Fisher and Martin Gervais WP-09-01 Why do the Elderly Save? The Role of Medical Expenses Mariacristina De Nardi, Eric French, and John Bailey Jones WP-09-02 Using Stock Returns to Identify Government Spending Shocks Jonas D.M. Fisher and Ryan Peters WP-09-03 Stochastic Volatility Torben G. Andersen and Luca Benzoni WP-09-04 The Effect of Disability Insurance Receipt on Labor Supply Eric French and Jae Song WP-09-05 CEO Overconfidence and Dividend Policy Sanjay Deshmukh, Anand M. Goel, and Keith M. Howe WP-09-06 Do Financial Counseling Mandates Improve Mortgage Choice and Performance? Evidence from a Legislative Experiment Sumit Agarwal,Gene Amromin, Itzhak Ben-David, Souphala Chomsisengphet, and Douglas D. Evanoff WP-09-07 Perverse Incentives at the Banks? Evidence from a Natural Experiment Sumit Agarwal and Faye H. Wang WP-09-08 Pay for Percentile Gadi Barlevy and Derek Neal WP-09-09 The Life and Times of Nicolas Dutot François R. Velde WP-09-10 Regulating Two-Sided Markets: An Empirical Investigation Santiago Carbó Valverde, Sujit Chakravorti, and Francisco Rodriguez Fernandez WP-09-11 2 Working Paper Series (continued) The Case of the Undying Debt François R. Velde Paying for Performance: The Education Impacts of a Community College Scholarship Program for Low-income Adults Lisa Barrow, Lashawn Richburg-Hayes, Cecilia Elena Rouse, and Thomas Brock Establishments Dynamics, Vacancies and Unemployment: A Neoclassical Synthesis Marcelo Veracierto WP-09-12 WP-09-13 WP-09-14 The Price of Gasoline and the Demand for Fuel Economy: Evidence from Monthly New Vehicles Sales Data Thomas Klier and Joshua Linn WP-09-15 Estimation of a Transformation Model with Truncation, Interval Observation and Time-Varying Covariates Bo E. Honoré and Luojia Hu WP-09-16 Self-Enforcing Trade Agreements: Evidence from Antidumping Policy Chad P. Bown and Meredith A. Crowley WP-09-17 Too much right can make a wrong: Setting the stage for the financial crisis Richard J. Rosen WP-09-18 Can Structural Small Open Economy Models Account for the Influence of Foreign Disturbances? Alejandro Justiniano and Bruce Preston WP-09-19 Liquidity Constraints of the Middle Class Jeffrey R. Campbell and Zvi Hercowitz WP-09-20 Monetary Policy and Uncertainty in an Empirical Small Open Economy Model Alejandro Justiniano and Bruce Preston WP-09-21 Firm boundaries and buyer-supplier match in market transaction: IT system procurement of U.S. credit unions Yukako Ono and Junichi Suzuki Health and the Savings of Insured Versus Uninsured, Working-Age Households in the U.S. Maude Toussaint-Comeau and Jonathan Hartley WP-09-22 WP-09-23 The Economics of “Radiator Springs:” Industry Dynamics, Sunk Costs, and Spatial Demand Shifts Jeffrey R. Campbell and Thomas N. Hubbard WP-09-24 On the Relationship between Mobility, Population Growth, and Capital Spending in the United States Marco Bassetto and Leslie McGranahan WP-09-25 The Impact of Rosenwald Schools on Black Achievement Daniel Aaronson and Bhashkar Mazumder WP-09-26 3 Working Paper Series (continued) Comment on “Letting Different Views about Business Cycles Compete” Jonas D.M. Fisher WP-10-01 Macroeconomic Implications of Agglomeration Morris A. Davis, Jonas D.M. Fisher and Toni M. Whited WP-10-02 Accounting for non-annuitization Svetlana Pashchenko WP-10-03 Robustness and Macroeconomic Policy Gadi Barlevy WP-10-04 Benefits of Relationship Banking: Evidence from Consumer Credit Markets Sumit Agarwal, Souphala Chomsisengphet, Chunlin Liu, and Nicholas S. Souleles WP-10-05 The Effect of Sales Tax Holidays on Household Consumption Patterns Nathan Marwell and Leslie McGranahan WP-10-06 Gathering Insights on the Forest from the Trees: A New Metric for Financial Conditions Scott Brave and R. Andrew Butters WP-10-07 Identification of Models of the Labor Market Eric French and Christopher Taber WP-10-08 Public Pensions and Labor Supply Over the Life Cycle Eric French and John Jones WP-10-09 Explaining Asset Pricing Puzzles Associated with the 1987 Market Crash Luca Benzoni, Pierre Collin-Dufresne, and Robert S. Goldstein WP-10-10 Prenatal Sex Selection and Girls’ Well‐Being: Evidence from India Luojia Hu and Analía Schlosser WP-10-11 Mortgage Choices and Housing Speculation Gadi Barlevy and Jonas D.M. Fisher WP-10-12 Did Adhering to the Gold Standard Reduce the Cost of Capital? Ron Alquist and Benjamin Chabot WP-10-13 Introduction to the Macroeconomic Dynamics: Special issues on money, credit, and liquidity Ed Nosal, Christopher Waller, and Randall Wright WP-10-14 Summer Workshop on Money, Banking, Payments and Finance: An Overview Ed Nosal and Randall Wright WP-10-15 Cognitive Abilities and Household Financial Decision Making Sumit Agarwal and Bhashkar Mazumder WP-10-16 Complex Mortgages Gene Amromin, Jennifer Huang, Clemens Sialm, and Edward Zhong WP-10-17 4 Working Paper Series (continued) The Role of Housing in Labor Reallocation Morris Davis, Jonas Fisher, and Marcelo Veracierto WP-10-18 Why Do Banks Reward their Customers to Use their Credit Cards? Sumit Agarwal, Sujit Chakravorti, and Anna Lunn WP-10-19 The impact of the originate-to-distribute model on banks before and during the financial crisis Richard J. Rosen WP-10-20 Simple Markov-Perfect Industry Dynamics Jaap H. Abbring, Jeffrey R. Campbell, and Nan Yang WP-10-21 Commodity Money with Frequent Search Ezra Oberfield and Nicholas Trachter WP-10-22 Corporate Average Fuel Economy Standards and the Market for New Vehicles Thomas Klier and Joshua Linn WP-11-01 The Role of Securitization in Mortgage Renegotiation Sumit Agarwal, Gene Amromin, Itzhak Ben-David, Souphala Chomsisengphet, and Douglas D. Evanoff WP-11-02 Market-Based Loss Mitigation Practices for Troubled Mortgages Following the Financial Crisis Sumit Agarwal, Gene Amromin, Itzhak Ben-David, Souphala Chomsisengphet, and Douglas D. Evanoff WP-11-03 Federal Reserve Policies and Financial Market Conditions During the Crisis Scott A. Brave and Hesna Genay WP-11-04 The Financial Labor Supply Accelerator Jeffrey R. Campbell and Zvi Hercowitz WP-11-05 Survival and long-run dynamics with heterogeneous beliefs under recursive preferences Jaroslav Borovička WP-11-06 A Leverage-based Model of Speculative Bubbles (Revised) Gadi Barlevy WP-11-07 Estimation of Panel Data Regression Models with Two-Sided Censoring or Truncation Sule Alan, Bo E. Honoré, Luojia Hu, and Søren Leth–Petersen WP-11-08 5