Estimating and Testing a Quantile Regression Model with Interactive Effects Matthew Harding1 and Carlos Lamarche2 1 Stanford University 2 University of Oklahoma California Econometrics Conference, Sept 24, 2010 Estimating and Testing a Quantile Regression Model with Interactive Effects 1 / 50 Motivation Motivation Classical least squares methods for panel data are often inadequate for empirical analysis. They deal with individual heterogeneity, but fail to estimate effects other than the mean. Koenker (2004), Lamarche (2010), Harding and Lamarche (2009), Abrevaya and Dahl (2008) suggest approaches but their use is limited under general conditions. Limitation They assume that latent heterogeneity has the classical additively separable, time-invariant structure. Estimating and Testing a Quantile Regression Model with Interactive Effects 2 / 50 Motivation Motivation Classical least squares methods for panel data are often inadequate for empirical analysis. They deal with individual heterogeneity, but fail to estimate effects other than the mean. Koenker (2004), Lamarche (2010), Harding and Lamarche (2009), Abrevaya and Dahl (2008) suggest approaches but their use is limited under general conditions. Limitation They assume that latent heterogeneity has the classical additively separable, time-invariant structure. Estimating and Testing a Quantile Regression Model with Interactive Effects 2 / 50 Motivation Motivation Classical least squares methods for panel data are often inadequate for empirical analysis. They deal with individual heterogeneity, but fail to estimate effects other than the mean. Koenker (2004), Lamarche (2010), Harding and Lamarche (2009), Abrevaya and Dahl (2008) suggest approaches but their use is limited under general conditions. Limitation They assume that latent heterogeneity has the classical additively separable, time-invariant structure. Estimating and Testing a Quantile Regression Model with Interactive Effects 2 / 50 Motivation Motivation Classical least squares methods for panel data are often inadequate for empirical analysis. They deal with individual heterogeneity, but fail to estimate effects other than the mean. Koenker (2004), Lamarche (2010), Harding and Lamarche (2009), Abrevaya and Dahl (2008) suggest approaches but their use is limited under general conditions. Limitation They assume that latent heterogeneity has the classical additively separable, time-invariant structure. Estimating and Testing a Quantile Regression Model with Interactive Effects 2 / 50 Motivation Motivation Classical least squares methods for panel data are often inadequate for empirical analysis. They deal with individual heterogeneity, but fail to estimate effects other than the mean. Koenker (2004), Lamarche (2010), Harding and Lamarche (2009), Abrevaya and Dahl (2008) suggest approaches but their use is limited under general conditions. Limitation The estimation of N nuisance parameters is computationally demanding. Estimating and Testing a Quantile Regression Model with Interactive Effects 3 / 50 Motivation Motivation Classical least squares methods for panel data are often inadequate for empirical analysis. They deal with individual heterogeneity, but fail to estimate effects other than the mean. Koenker (2004), Lamarche (2010), Harding and Lamarche (2009), Abrevaya and Dahl (2008) suggest approaches but their use is limited under general conditions. Contribution This paper offers a simple procedure that allows estimation of distributional effects, under mild conditions. Estimating and Testing a Quantile Regression Model with Interactive Effects 4 / 50 Motivation Motivation Classical least squares methods for panel data are often inadequate for empirical analysis. They deal with individual heterogeneity, but fail to estimate effects other than the mean. Koenker (2004), Lamarche (2010), Harding and Lamarche (2009), Abrevaya and Dahl (2008) suggest approaches but their use is limited under general conditions. Contribution This paper offers a simple procedure that allows estimation of distributional effects, under mild conditions. Estimating and Testing a Quantile Regression Model with Interactive Effects 4 / 50 Motivation Motivation Classical least squares methods for panel data are often inadequate for empirical analysis. They deal with individual heterogeneity, but fail to estimate effects other than the mean. Koenker (2004), Lamarche (2010), Harding and Lamarche (2009), Abrevaya and Dahl (2008) suggest approaches but their use is limited under general conditions. Contribution This paper offers a simple procedure that allows estimation of distributional effects, under mild conditions. Estimating and Testing a Quantile Regression Model with Interactive Effects 4 / 50 Background In the last half a century, understanding the drivers of students’ academic performance has been a major focus in the economics of education. A number of studies have focused on class size and peer effects (e.g., Coleman 1966, Krueger 1999, Hoxby 2000, Hanushek et al. 2003). The empirical evidence on the effect of class size and class composition on achievement remains mixed. The literature offers a number of studies on the mean effect, but few studies investigate its distributional effect. One exception is Ma and Koenker (2006). Estimating and Testing a Quantile Regression Model with Interactive Effects 5 / 50 Background In the last half a century, understanding the drivers of students’ academic performance has been a major focus in the economics of education. A number of studies have focused on class size and peer effects (e.g., Coleman 1966, Krueger 1999, Hoxby 2000, Hanushek et al. 2003). The empirical evidence on the effect of class size and class composition on achievement remains mixed. The literature offers a number of studies on the mean effect, but few studies investigate its distributional effect. One exception is Ma and Koenker (2006). Estimating and Testing a Quantile Regression Model with Interactive Effects 5 / 50 Background In the last half a century, understanding the drivers of students’ academic performance has been a major focus in the economics of education. A number of studies have focused on class size and peer effects (e.g., Coleman 1966, Krueger 1999, Hoxby 2000, Hanushek et al. 2003). The empirical evidence on the effect of class size and class composition on achievement remains mixed. The literature offers a number of studies on the mean effect, but few studies investigate its distributional effect. One exception is Ma and Koenker (2006). Estimating and Testing a Quantile Regression Model with Interactive Effects 5 / 50 Background In the last half a century, understanding the drivers of students’ academic performance has been a major focus in the economics of education. A number of studies have focused on class size and peer effects (e.g., Coleman 1966, Krueger 1999, Hoxby 2000, Hanushek et al. 2003). The empirical evidence on the effect of class size and class composition on achievement remains mixed. The literature offers a number of studies on the mean effect, but few studies investigate its distributional effect. One exception is Ma and Koenker (2006). Estimating and Testing a Quantile Regression Model with Interactive Effects 5 / 50 20 22 24 test scores 26 28 30 Background −0.5 0.0 0.5 1.0 1.5 large class Estimating and Testing a Quantile Regression Model with Interactive Effects 6 / 50 test scores 0.10 0.05 35 30 25 20 0.00 f(y|x) 0.15 0.20 Background −0.5 15 0.0 0.5 1.0 1.5 large class Estimating and Testing a Quantile Regression Model with Interactive Effects 7 / 50 test scores 26 28 30 Background 20 22 24 ^ !1=−0.16 −0.5 0.0 0.5 1.0 1.5 large class Estimating and Testing a Quantile Regression Model with Interactive Effects 8 / 50 test scores 26 28 30 Background 22 24 ^ !1=−0.16 20 ^ !1(0.1)=−0.30 −0.5 0.0 0.5 1.0 1.5 large class Estimating and Testing a Quantile Regression Model with Interactive Effects 9 / 50 30 Background test scores 26 28 ^ !1(0.9)=−0.08 22 24 ^ !1=−0.16 20 ^ !1(0.1)=−0.30 −0.5 0.0 0.5 1.0 1.5 large class Estimating and Testing a Quantile Regression Model with Interactive Effects 10 / 50 Background It is standard to consider (e.g., Hanushek et al. 2003): yict = d 0ct α + x 0i β + λi + Fct + uict The λi ’s are associated with motivation and ability, and the Fct ’s measure teaching quality. Note that we are imposing λi + Fct . High teaching quality may have a modest effect on performance among unmotivated students, while it may dramatically affect strong, motivated students. Can we estimate this model? Estimating and Testing a Quantile Regression Model with Interactive Effects 11 / 50 Background It is standard to consider (e.g., Hanushek et al. 2003): yict = d 0ct α + x 0i β + λi + Fct + uict The λi ’s are associated with motivation and ability, and the Fct ’s measure teaching quality. Note that we are imposing λi + Fct . High teaching quality may have a modest effect on performance among unmotivated students, while it may dramatically affect strong, motivated students. Can we estimate this model? Estimating and Testing a Quantile Regression Model with Interactive Effects 11 / 50 Background It is standard to consider (e.g., Hanushek et al. 2003): yict = d 0ct α + x 0i β + λi + Fct + uict The λi ’s are associated with motivation and ability, and the Fct ’s measure teaching quality. Note that we are imposing λi + Fct . High teaching quality may have a modest effect on performance among unmotivated students, while it may dramatically affect strong, motivated students. Can we estimate this model? Estimating and Testing a Quantile Regression Model with Interactive Effects 11 / 50 Background It is standard to consider (e.g., Hanushek et al. 2003): yict = d 0ct α + x 0i β + λi + Fct + uict The λi ’s are associated with motivation and ability, and the Fct ’s measure teaching quality. Note that we are imposing λi + Fct . High teaching quality may have a modest effect on performance among unmotivated students, while it may dramatically affect strong, motivated students. Can we estimate this model? Estimating and Testing a Quantile Regression Model with Interactive Effects 11 / 50 Background It is standard to consider (e.g., Hanushek et al. 2003): yict = d 0ct α + x 0i β + λi + Fct + uict The λi ’s are associated with motivation and ability, and the Fct ’s measure teaching quality. Note that we are imposing λi + Fct . High teaching quality may have a modest effect on performance among unmotivated students, while it may dramatically affect strong, motivated students. Can we estimate this model? Estimating and Testing a Quantile Regression Model with Interactive Effects 11 / 50 Models and Estimators Asymptotic Theory Application Conclusions Outline Estimating and Testing a Quantile Regression Model with Interactive Effects 12 / 50 Models and Estimators Asymptotic Theory Application Conclusions A Panel Data Model Recently, Pesaran (2006) and Bai (2009) write, yit = α0 d it + β 0 x it + λ0i F t + uit . A panel data model with r factors, λ0i F t = λi1 Ft1 + λi2 Ft2 + . . . + λir Ftr . If r = Ft = 1, model with individual effects: λi . If r = λi = 1, model with time effects: Ft . If r = 2 and λi2 = Ft1 = 1, model with additive individual and time effects: λi + Ft . Estimating and Testing a Quantile Regression Model with Interactive Effects 13 / 50 Models and Estimators Asymptotic Theory Application Conclusions A Panel Data Model Recently, Pesaran (2006) and Bai (2009) write, yit = α0 d it + β 0 x it + λ0i F t + uit . A panel data model with r factors, λ0i F t = λi1 Ft1 + λi2 Ft2 + . . . + λir Ftr . If r = Ft = 1, model with individual effects: λi . If r = λi = 1, model with time effects: Ft . If r = 2 and λi2 = Ft1 = 1, model with additive individual and time effects: λi + Ft . Estimating and Testing a Quantile Regression Model with Interactive Effects 13 / 50 Models and Estimators Asymptotic Theory Application Conclusions A Panel Data Model Recently, Pesaran (2006) and Bai (2009) write, yit = α0 d it + β 0 x it + λ0i F t + uit . A panel data model with r factors, λ0i F t = λi1 Ft1 + λi2 Ft2 + . . . + λir Ftr . If r = Ft = 1, model with individual effects: λi . If r = λi = 1, model with time effects: Ft . If r = 2 and λi2 = Ft1 = 1, model with additive individual and time effects: λi + Ft . Estimating and Testing a Quantile Regression Model with Interactive Effects 13 / 50 Models and Estimators Asymptotic Theory Application Conclusions A Panel Data Model Recently, Pesaran (2006) and Bai (2009) write, yit = α0 d it + β 0 x it + λ0i F t + uit . A panel data model with r factors, λ0i F t = λi1 Ft1 + λi2 Ft2 + . . . + λir Ftr . If r = Ft = 1, model with individual effects: λi . If r = λi = 1, model with time effects: Ft . If r = 2 and λi2 = Ft1 = 1, model with additive individual and time effects: λi + Ft . Estimating and Testing a Quantile Regression Model with Interactive Effects 13 / 50 Models and Estimators Asymptotic Theory Application Conclusions A Panel Data Model Recently, Pesaran (2006) and Bai (2009) write, yit = α0 d it + β 0 x it + λ0i F t + uit . A panel data model with r factors, λ0i F t = λi1 Ft1 + λi2 Ft2 + . . . + λir Ftr . If r = Ft = 1, model with individual effects: λi . If r = λi = 1, model with time effects: Ft . If r = 2 and λi2 = Ft1 = 1, model with additive individual and time effects: λi + Ft . Estimating and Testing a Quantile Regression Model with Interactive Effects 13 / 50 Models and Estimators Asymptotic Theory Application Conclusions A Panel Data Model Recently, Pesaran (2006) and Bai (2009) write, yit = α0 d it + β 0 x it + λ0i F t + uit . A panel data model with r factors, λ0i F t = λi1 Ft1 + λi2 Ft2 + . . . + λir Ftr . The time effect Ft is stochastically dependent on d it . The effects (Ft , λi ) are stochastically dependent on d it . The variables (Ft , λi , uit ) and d it are stochastically dependent. Estimating and Testing a Quantile Regression Model with Interactive Effects 14 / 50 Models and Estimators Asymptotic Theory Application Conclusions A Panel Data Model Recently, Pesaran (2006) and Bai (2009) write, yit = α0 d it + β 0 x it + λ0i F t + uit . A panel data model with r factors, λ0i F t = λi1 Ft1 + λi2 Ft2 + . . . + λir Ftr . The time effect Ft is stochastically dependent on d it . The effects (Ft , λi ) are stochastically dependent on d it . The variables (Ft , λi , uit ) and d it are stochastically dependent. Estimating and Testing a Quantile Regression Model with Interactive Effects 14 / 50 Models and Estimators Asymptotic Theory Application Conclusions A Panel Data Model Recently, Pesaran (2006) and Bai (2009) write, yit = α0 d it + β 0 x it + λ0i F t + uit . A panel data model with r factors, λ0i F t = λi1 Ft1 + λi2 Ft2 + . . . + λir Ftr . The time effect Ft is stochastically dependent on d it . The effects (Ft , λi ) are stochastically dependent on d it . The variables (Ft , λi , uit ) and d it are stochastically dependent. Estimating and Testing a Quantile Regression Model with Interactive Effects 14 / 50 Models and Estimators Asymptotic Theory Application Conclusions A Panel Data Model Recently, Pesaran (2006) and Bai (2009) write, yit = α0 d it + β 0 x it + λ0i F t + uit . A panel data model with r factors, λ0i F t = λi1 Ft1 + λi2 Ft2 + . . . + λir Ftr . The time effect Ft is stochastically dependent on d it . The effects (Ft , λi ) are stochastically dependent on d it . The variables (Ft , λi , uit ) and d it are stochastically dependent. Estimating and Testing a Quantile Regression Model with Interactive Effects 14 / 50 Models and Estimators Asymptotic Theory Application Conclusions A Panel Data Model Recently, Pesaran (2006) and Bai (2009) write, yit = α0 d it + β 0 x it + λ0i F t + uit . A panel data model with r factors, λ0i F t = λi1 Ft1 + λi2 Ft2 + . . . + λir Ftr . The time effect Ft is stochastically dependent on d it . The effects (Ft , λi ) are stochastically dependent on d it . The variables (Ft , λi , uit ) and d it are stochastically dependent. Estimating and Testing a Quantile Regression Model with Interactive Effects 14 / 50 Models and Estimators Asymptotic Theory Application Conclusions Least Squares Estimation of a Panel Data Model Lemma Under regularity conditions, α can be estimated by α̂ = (D 0 P̄ M̄W D)−1 (D 0 P̄ M̄W y) where P̄ M̄W is a projection matrix that uses instruments W and cross-sectional averages. Remark The method extends Pesaran (2006) analysis accommodating to issues associated with dependence between d and (λ0 , u)0 . Remark It is possible to obtain a feasible estimator in a model with interactive effects and endogenous covariates. Estimating and Testing a Quantile Regression Model with Interactive Effects 15 / 50 Models and Estimators Asymptotic Theory Application Conclusions Least Squares Estimation of a Panel Data Model Lemma Under regularity conditions, α can be estimated by α̂ = (D 0 P̄ M̄W D)−1 (D 0 P̄ M̄W y) where P̄ M̄W is a projection matrix that uses instruments W and cross-sectional averages. Remark The method extends Pesaran (2006) analysis accommodating to issues associated with dependence between d and (λ0 , u)0 . Remark It is possible to obtain a feasible estimator in a model with interactive effects and endogenous covariates. Estimating and Testing a Quantile Regression Model with Interactive Effects 15 / 50 Models and Estimators Asymptotic Theory Application Conclusions Least Squares Estimation of a Panel Data Model Lemma Under regularity conditions, α can be estimated by α̂ = (D 0 P̄ M̄W D)−1 (D 0 P̄ M̄W y) where P̄ M̄W is a projection matrix that uses instruments W and cross-sectional averages. Remark The method extends Pesaran (2006) analysis accommodating to issues associated with dependence between d and (λ0 , u)0 . Remark It is possible to obtain a feasible estimator in a model with interactive effects and endogenous covariates. Estimating and Testing a Quantile Regression Model with Interactive Effects 15 / 50 Models and Estimators Asymptotic Theory Application Conclusions The Quantile Regression Model The paper considers conditional quantile functions of the form QYit (τ |d it , x it , λi , F t ) = α(τ )0 d it + β(τ )0 x it + λi (τ )0 F t (τ ) where τj ∈ (0, 1) is the quantile of interest, and QYit (τj |d it , x it , λi , F t ) ≡ inf{yit : FYit (yit |d it , x it , λi , F t ) ≥ τj } The covariate’s effect is to shift the location, scale and possibly shape of the conditional distribution of the response. The model also allows for individual and time specific distributional shifts. Estimating and Testing a Quantile Regression Model with Interactive Effects 16 / 50 Models and Estimators Asymptotic Theory Application Conclusions The Quantile Regression Model The paper considers conditional quantile functions of the form QYit (τ |d it , x it , λi , F t ) = α(τ )0 d it + β(τ )0 x it + λi (τ )0 F t (τ ) where τj ∈ (0, 1) is the quantile of interest, and QYit (τj |d it , x it , λi , F t ) ≡ inf{yit : FYit (yit |d it , x it , λi , F t ) ≥ τj } The covariate’s effect is to shift the location, scale and possibly shape of the conditional distribution of the response. The model also allows for individual and time specific distributional shifts. Estimating and Testing a Quantile Regression Model with Interactive Effects 16 / 50 Models and Estimators Asymptotic Theory Application Conclusions The Quantile Regression Model The paper considers conditional quantile functions of the form QYit (τ |d it , x it , λi , F t ) = α(τ )0 d it + β(τ )0 x it + λi (τ )0 F t (τ ) where τj ∈ (0, 1) is the quantile of interest, and QYit (τj |d it , x it , λi , F t ) ≡ inf{yit : FYit (yit |d it , x it , λi , F t ) ≥ τj } The covariate’s effect is to shift the location, scale and possibly shape of the conditional distribution of the response. The model also allows for individual and time specific distributional shifts. Estimating and Testing a Quantile Regression Model with Interactive Effects 16 / 50 Models and Estimators Asymptotic Theory Application Conclusions An Estimator for Panel Data The estimator θ̂(τ ) ≡ α̂(τ ), β̂(α̂(τ ), τ )), δ̂(α̂(τ ), τ )) is arg min β,γ,δ T X N X ρτ (yit − d 0it α − x 0it β − Ψ̂0t (τ )δ − Φ̂0it (τ )γ), t=1 i=1 where ρτ is the quantile regression check function and, α̂(τ ) = arg min γ̂(τ, α)0 Aγ̂(τ, α) α The first term provides an asymptotic (consistent) approximation for the interactive effect. The second term is a vector of transformations of instruments, as in the classical IV case. Estimating and Testing a Quantile Regression Model with Interactive Effects 17 / 50 Models and Estimators Asymptotic Theory Application Conclusions An Estimator for Panel Data The estimator θ̂(τ ) ≡ α̂(τ ), β̂(α̂(τ ), τ )), δ̂(α̂(τ ), τ )) is arg min β,γ,δ T X N X ρτ (yit − d 0it α − x 0it β − Ψ̂0t (τ )δ − Φ̂0it (τ )γ), t=1 i=1 where ρτ is the quantile regression check function and, α̂(τ ) = arg min γ̂(τ, α)0 Aγ̂(τ, α) α The first term provides an asymptotic (consistent) approximation for the interactive effect. The second term is a vector of transformations of instruments, as in the classical IV case. Estimating and Testing a Quantile Regression Model with Interactive Effects 17 / 50 Models and Estimators Asymptotic Theory Application Conclusions An Estimator for Panel Data The estimator θ̂(τ ) ≡ α̂(τ ), β̂(α̂(τ ), τ )), δ̂(α̂(τ ), τ )) is arg min β,γ,δ T X N X ρτ (yit − d 0it α − x 0it β − Ψ̂0t (τ )δ − Φ̂0it (τ )γ), t=1 i=1 where ρτ is the quantile regression check function and, α̂(τ ) = arg min γ̂(τ, α)0 Aγ̂(τ, α) α The first term provides an asymptotic (consistent) approximation for the interactive effect. The second term is a vector of transformations of instruments, as in the classical IV case. Estimating and Testing a Quantile Regression Model with Interactive Effects 17 / 50 Models and Estimators Asymptotic Theory Application Conclusions Existing Fixed Effects Methods The estimator considered in Koenker (2004), arg min α,β,λ T X N X ρτ (yit − d 0it α − x 0it β − λi ) t=1 i=1 Our method is similar to Harding and Lamarche (2009), but arg min β,γ T X N X ρτ (yit − d 0it α − x 0it β − λi − Φ̂0it (τ )γ) t=1 i=1 Estimation of N nuisance parameters could be, in some applications, computationally demanding. They can produce biased results under mild conditions. Estimating and Testing a Quantile Regression Model with Interactive Effects 18 / 50 Models and Estimators Asymptotic Theory Application Conclusions Existing Fixed Effects Methods The estimator considered in Koenker (2004), arg min α,β,λ T X N X ρτ (yit − d 0it α − x 0it β − λi ) t=1 i=1 Our method is similar to Harding and Lamarche (2009), but arg min β,γ T X N X ρτ (yit − d 0it α − x 0it β − λi − Φ̂0it (τ )γ) t=1 i=1 Estimation of N nuisance parameters could be, in some applications, computationally demanding. They can produce biased results under mild conditions. Estimating and Testing a Quantile Regression Model with Interactive Effects 18 / 50 Models and Estimators Asymptotic Theory Application Conclusions Existing Fixed Effects Methods The estimator considered in Koenker (2004), arg min α,β,λ T X N X ρτ (yit − d 0it α − x 0it β − λi ) t=1 i=1 Our method is similar to Harding and Lamarche (2009), but arg min β,γ T X N X ρτ (yit − d 0it α − x 0it β − λi − Φ̂0it (τ )γ) t=1 i=1 Estimation of N nuisance parameters could be, in some applications, computationally demanding. They can produce biased results under mild conditions. Estimating and Testing a Quantile Regression Model with Interactive Effects 18 / 50 Models and Estimators Asymptotic Theory Application Conclusions Existing Fixed Effects Methods The estimator considered in Koenker (2004), arg min α,β,λ T X N X ρτ (yit − d 0it α − x 0it β − λi ) t=1 i=1 Our method is similar to Harding and Lamarche (2009), but arg min β,γ T X N X ρτ (yit − d 0it α − x 0it β − λi − Φ̂0it (τ )γ) t=1 i=1 Estimation of N nuisance parameters could be, in some applications, computationally demanding. They can produce biased results under mild conditions. Estimating and Testing a Quantile Regression Model with Interactive Effects 18 / 50 Models and Estimators Asymptotic Theory Application Conclusions Useful variation on these models Combine fixed effects with interactive effects specification arg min β,γ T X N X ρτ (yit − d 0it α − x 0it β − λi − Ψ̂0it (τ )δ − Φ̂0it (τ )γ) t=1 i=1 Can be used to test whether the interactive effects are useful Estimating and Testing a Quantile Regression Model with Interactive Effects 19 / 50 Models and Estimators Asymptotic Theory Application Conclusions Regularity Conditions Regularity Conditions 1 Yit has a conditional distribution Fit , and continuous densities fit bounded away from 0 and ∞ at ξit (τj ). 2 (α(τ ), β(τ ), δ(τ )) ∈ int of a compact and convex set. 3 √ max kz it k/ NT → 0, for z = {d, x, w }. 4 The Jacobian matrices have full rank and are continuous. 5 There exist limiting positive definite matrices S(τ ) and J(τ ). Estimating and Testing a Quantile Regression Model with Interactive Effects 20 / 50 Models and Estimators Asymptotic Theory Application Conclusions Regularity Conditions Regularity Conditions 1 Yit has a conditional distribution Fit , and continuous densities fit bounded away from 0 and ∞ at ξit (τj ). 2 (α(τ ), β(τ ), δ(τ )) ∈ int of a compact and convex set. 3 √ max kz it k/ NT → 0, for z = {d, x, w }. 4 The Jacobian matrices have full rank and are continuous. 5 There exist limiting positive definite matrices S(τ ) and J(τ ). Estimating and Testing a Quantile Regression Model with Interactive Effects 20 / 50 Models and Estimators Asymptotic Theory Application Conclusions Regularity Conditions Regularity Conditions 1 Yit has a conditional distribution Fit , and continuous densities fit bounded away from 0 and ∞ at ξit (τj ). 2 (α(τ ), β(τ ), δ(τ )) ∈ int of a compact and convex set. 3 √ max kz it k/ NT → 0, for z = {d, x, w }. 4 The Jacobian matrices have full rank and are continuous. 5 There exist limiting positive definite matrices S(τ ) and J(τ ). Estimating and Testing a Quantile Regression Model with Interactive Effects 20 / 50 Models and Estimators Asymptotic Theory Application Conclusions Regularity Conditions Regularity Conditions 1 Yit has a conditional distribution Fit , and continuous densities fit bounded away from 0 and ∞ at ξit (τj ). 2 (α(τ ), β(τ ), δ(τ )) ∈ int of a compact and convex set. 3 √ max kz it k/ NT → 0, for z = {d, x, w }. 4 The Jacobian matrices have full rank and are continuous. 5 There exist limiting positive definite matrices S(τ ) and J(τ ). Estimating and Testing a Quantile Regression Model with Interactive Effects 20 / 50 Models and Estimators Asymptotic Theory Application Conclusions Regularity Conditions Regularity Conditions 1 Yit has a conditional distribution Fit , and continuous densities fit bounded away from 0 and ∞ at ξit (τj ). 2 (α(τ ), β(τ ), δ(τ )) ∈ int of a compact and convex set. 3 √ max kz it k/ NT → 0, for z = {d, x, w }. 4 The Jacobian matrices have full rank and are continuous. 5 There exist limiting positive definite matrices S(τ ) and J(τ ). Estimating and Testing a Quantile Regression Model with Interactive Effects 20 / 50 Models and Estimators Asymptotic Theory Application Conclusions Theoretical Results Theorem Under the regularity conditions, the estimator (α̂(τ )0 , β̂(τ )0 )0 is consistent and asymptotically normally distributed with mean (α(τ )0 , β(τ )0 )0 and covariance matrix J 0 (τ )S(τ )J(τ ). Remark The paper suggests ways of doing inference. The standard errors are obtained by estimating the asymptotic covariance matrices. Estimating and Testing a Quantile Regression Model with Interactive Effects 21 / 50 Models and Estimators Asymptotic Theory Application Conclusions Theoretical Results Theorem Under the regularity conditions, the estimator (α̂(τ )0 , β̂(τ )0 )0 is consistent and asymptotically normally distributed with mean (α(τ )0 , β(τ )0 )0 and covariance matrix J 0 (τ )S(τ )J(τ ). Remark The paper suggests ways of doing inference. The standard errors are obtained by estimating the asymptotic covariance matrices. Estimating and Testing a Quantile Regression Model with Interactive Effects 21 / 50 Models and Estimators Asymptotic Theory Application Conclusions Simulation design: yit = β0 + β1 dit + γxt + λ1i f1t + λ2i f2t + (1 + hdit )uit dit = π0 + π1 wit + π2 xt + π3 f1t + π3 f2t + π4 λ1i f1t + π4 λ2i f2t + i + vit fjt = ρf fjt−1 + ηjt ηjt = ρη ηjt−1 + ejt for j = {1, 2}, . . . t = −49, . . . 0, . . . T in the last two equations. The random variables are xt ∼ N (0, 1), λi1 , λi2 ∼ N (1, 0.2), and e, and w are Gaussian independent random variables. The error terms are (uit , vit )0 ∼ (0, Ω), distributed either as Gaussian or t-student distribution with two degrees of freedom. The parameters are assumed to be: β0 = π3 = 2, β1 = γ = π0 = π1 = π2 = 1, ρf = 0.90, ρη = 0.25, and Ω11 = Ω22 = 1. Estimating and Testing a Quantile Regression Model with Interactive Effects 22 / 50 Models and Estimators Asymptotic Theory Application Conclusions Design 1 The endogenous variable d is not correlated with the λ’s, and the variables u and v are independent Gaussian variables. Although d is not correlated with the individual effects and the error term, it is correlated with the F ’s. We assume π4 = 0 and Ω12 = Ω21 = 0. Design 2 The variable d is correlated with F ’s and λ’s, and the error terms in equations 2.1 and 2.1 are not correlated. We assume π4 = 2 and Ω12 = Ω21 = 0. Design 3 The error terms in equations 2.1 and 2.1 are now correlated, assuming that Ω12 = Ω21 = 0.5. The variable d is also correlated with the F ’s and λ’s as in the experiment carried out in Design 2. Estimating and Testing a Quantile Regression Model with Interactive Effects 23 / 50 Models and Estimators Asymptotic Theory Application Conclusions Monte Carlo Evidence (cont.) We consider the same model expanding the design to include different sample sizes N = {50, 100} and T = {4, 8}. We compare estimators: (1) quantile regression estimator (QR); (2) Koenker’s (2004) fixed effects estimator (QRFE); (3) Harding and Lamarche’s (2009) instrumental variable estimator (QRIVFE); (4) the proposed approach (QRIE). We report bias and root mean square error (RMSE) on 24 different monte carlo experiments. Estimating and Testing a Quantile Regression Model with Interactive Effects 24 / 50 Models and Estimators Asymptotic Theory Application Conclusions Monte Carlo Evidence (cont.) We consider the same model expanding the design to include different sample sizes N = {50, 100} and T = {4, 8}. We compare estimators: (1) quantile regression estimator (QR); (2) Koenker’s (2004) fixed effects estimator (QRFE); (3) Harding and Lamarche’s (2009) instrumental variable estimator (QRIVFE); (4) the proposed approach (QRIE). We report bias and root mean square error (RMSE) on 24 different monte carlo experiments. Estimating and Testing a Quantile Regression Model with Interactive Effects 24 / 50 Models and Estimators Asymptotic Theory Application Conclusions Monte Carlo Evidence (cont.) We consider the same model expanding the design to include different sample sizes N = {50, 100} and T = {4, 8}. We compare estimators: (1) quantile regression estimator (QR); (2) Koenker’s (2004) fixed effects estimator (QRFE); (3) Harding and Lamarche’s (2009) instrumental variable estimator (QRIVFE); (4) the proposed approach (QRIE). We report bias and root mean square error (RMSE) on 24 different monte carlo experiments. Estimating and Testing a Quantile Regression Model with Interactive Effects 24 / 50 Models and Estimators Asymptotic Theory Application Conclusions Monte Carlo Evidence (cont.) We consider the same model expanding the design to include different sample sizes N = {50, 100} and T = {4, 8}. We compare estimators: (1) quantile regression estimator (QR); (2) Koenker’s (2004) fixed effects estimator (QRFE); (3) Harding and Lamarche’s (2009) instrumental variable estimator (QRIVFE); (4) the proposed approach (QRIE). We report bias and root mean square error (RMSE) on 24 different monte carlo experiments. Estimating and Testing a Quantile Regression Model with Interactive Effects 24 / 50 Models and Estimators Asymptotic Theory Application Conclusions Monte Carlo Evidence (cont.) We consider the same model expanding the design to include different sample sizes N = {50, 100} and T = {4, 8}. We compare estimators: (1) quantile regression estimator (QR); (2) Koenker’s (2004) fixed effects estimator (QRFE); (3) Harding and Lamarche’s (2009) instrumental variable estimator (QRIVFE); (4) the proposed approach (QRIE). We report bias and root mean square error (RMSE) on 24 different monte carlo experiments. Estimating and Testing a Quantile Regression Model with Interactive Effects 24 / 50 Models and Estimators Asymptotic Theory Application Conclusions Monte Carlo Evidence (cont.) We consider the same model expanding the design to include different sample sizes N = {50, 100} and T = {4, 8}. We compare estimators: (1) quantile regression estimator (QR); (2) Koenker’s (2004) fixed effects estimator (QRFE); (3) Harding and Lamarche’s (2009) instrumental variable estimator (QRIVFE); (4) the proposed approach (QRIE). We report bias and root mean square error (RMSE) on 24 different monte carlo experiments. Estimating and Testing a Quantile Regression Model with Interactive Effects 24 / 50 Models and Estimators Asymptotic Theory Application Conclusions Monte Carlo Evidence (cont.) We consider the same model expanding the design to include different sample sizes N = {50, 100} and T = {4, 8}. We compare estimators: (1) quantile regression estimator (QR); (2) Koenker’s (2004) fixed effects estimator (QRFE); (3) Harding and Lamarche’s (2009) instrumental variable estimator (QRIVFE); (4) the proposed approach (QRIE). We report bias and root mean square error (RMSE) on 24 different monte carlo experiments. Estimating and Testing a Quantile Regression Model with Interactive Effects 24 / 50 Models and Estimators Asymptotic Theory Application Conclusions 0.4 Comparing Methods 0.3 QR QRFE QRIVFE QRIE QR: D1, N=50, T=4, N(0,1) 0.1 Bias 0.2 QRFE: D1, N=50, T=4, N(0,1) 0.0 QRIVFE: D1, N=50, T=4, N(0,1) −0.1 QRIE: D1, N=50, T=4, N(0,1) 5 10 15 20 Monte Carlo experiment Estimating and Testing a Quantile Regression Model with Interactive Effects 25 / 50 Models and Estimators Asymptotic Theory Application Conclusions 0.4 Comparing Methods QR QRFE QRIVFE QRIE QR: D1, N=50, T=8, N(0,1) QRIVFE: D1, N=50, T=8, N(0,1) QRIE: D1, N=50, T=8, N(0,1) −0.1 0.0 0.1 Bias 0.2 0.3 QRFE: D1, N=50, T=8, N(0,1) 5 10 15 20 Monte Carlo experiment Estimating and Testing a Quantile Regression Model with Interactive Effects 26 / 50 Models and Estimators Asymptotic Theory Application Conclusions 0.4 Comparing Methods 0.3 QR QRFE QRIVFE QRIE Bias 0.2 QRFE: D1, N=100, T=4, N(0,1) QR: D1, N=100, T=4, N(0,1) 0.0 0.1 QRIVFE: D1, N=100, T=4, N(0,1) −0.1 QRIE: D1, N=100, T=4, N(0,1) 5 10 15 20 Monte Carlo experiment Estimating and Testing a Quantile Regression Model with Interactive Effects 27 / 50 Models and Estimators Asymptotic Theory Application Conclusions 0.4 Comparing Methods QR QRFE QRIVFE QRIE QRFE: D1, N=100, T=8, N(0,1) 0.1 Bias 0.2 0.3 QR: D1, N=100, T=8, N(0,1) 0.0 QRIVFE: D1, N=100, T=8, N(0,1) −0.1 QRIE: D1, N=100, T=8, N(0,1) 5 10 15 20 Monte Carlo experiment Estimating and Testing a Quantile Regression Model with Interactive Effects 28 / 50 Models and Estimators Asymptotic Theory Application Conclusions 0.4 Comparing Methods 0.3 QR QRFE QRIVFE QRIE QRFE: D1, N=50, T=4, t_2 0.1 Bias 0.2 QR: D1, N=50, T=4, t_2 0.0 QRIE: D1, N=50, T=4, t_2 −0.1 QRIVFE: D1, N=50, T=4, t_2 5 10 15 20 Monte Carlo experiment Estimating and Testing a Quantile Regression Model with Interactive Effects 29 / 50 Models and Estimators Asymptotic Theory Application Conclusions 0.4 Comparing Methods QR QRFE QRIVFE QRIE QRFE: D1, N=50, T=8, t_2 0.1 Bias 0.2 0.3 QR: D1, N=50, T=8, t_2 0.0 QRIE: D1, N=50, T=8, t_2 −0.1 QRIVFE: D1, N=50, T=8, t_2 5 10 15 20 Monte Carlo experiment Estimating and Testing a Quantile Regression Model with Interactive Effects 30 / 50 Models and Estimators Asymptotic Theory Application Conclusions 0.4 Comparing Methods 0.3 QR QRFE QRIVFE QRIE 0.2 QRFE: D1, N=100, T=4, t_2 0.1 Bias QR: D1, N=100, T=4, t_2 0.0 QRIE: D1, N=100, T=4, t_2 −0.1 QRIVFE: D1, N=100, T=4, t_2 5 10 15 20 Monte Carlo experiment Estimating and Testing a Quantile Regression Model with Interactive Effects 31 / 50 Models and Estimators Asymptotic Theory Application Conclusions 0.4 Comparing Methods QR QRFE QRIVFE QRIE QRFE: D1, N=100, T=8, t_2 0.1 Bias 0.2 0.3 QR: D1, N=100, T=8, t_2 0.0 QRIE: D1, N=100, T=8, t_2 −0.1 QRIVFE: D1, N=100, T=8, t_2 5 10 15 20 Monte Carlo experiment Estimating and Testing a Quantile Regression Model with Interactive Effects 32 / 50 Models and Estimators Asymptotic Theory Application Conclusions 0.4 Comparing Methods 0.3 QR QRFE QRIVFE QRIE 0.2 QR: D2, N=50, T=4, N(0,1) 0.1 Bias QRIVFE: D2, N=50, T=4, N(0,1) −0.1 0.0 QRIE: D2, N=50, T=4, N(0,1) 5 10 15 20 Monte Carlo experiment Estimating and Testing a Quantile Regression Model with Interactive Effects 33 / 50 Models and Estimators Asymptotic Theory Application Conclusions 0.4 Comparing Methods 0.3 QR QRFE QRIVFE QRIE QR: D2, N=50, T=8, N(0,1) 0.1 Bias 0.2 QRFE: D2, N=50, T=8, N(0,1) 0.0 QRIVFE: D2, N=50, T=8, N(0,1) −0.1 QRIE: D2, N=50, T=8, N(0,1) 5 10 15 20 Monte Carlo experiment Estimating and Testing a Quantile Regression Model with Interactive Effects 34 / 50 Models and Estimators Asymptotic Theory Application Conclusions 0.4 Comparing Methods 0.3 QR QRFE QRIVFE QRIE 0.2 QR: D2, N=100, T=4, N(0,1) 0.1 Bias QRFE: D2, N=100, T=4, N(0,1) 0.0 QRIVFE: D2, N=100, T=4, N(0,1) −0.1 QRIE: D2, N=100, T=4, N(0,1) 5 10 15 20 Monte Carlo experiment Estimating and Testing a Quantile Regression Model with Interactive Effects 35 / 50 Models and Estimators Asymptotic Theory Application Conclusions 0.4 Comparing Methods 0.3 QR QRFE QRIVFE QRIE QR: D2, N=100, T=8, N(0,1) 0.1 Bias 0.2 QRFE: D2, N=100, T=8, N(0,1) 0.0 QRIVFE: D2, N=100, T=8, N(0,1) −0.1 QRIE: D2, N=100, T=8, N(0,1) 5 10 15 20 Monte Carlo experiment Estimating and Testing a Quantile Regression Model with Interactive Effects 36 / 50 Models and Estimators Asymptotic Theory Application Conclusions 0.4 Comparing Methods 0.3 QR QRFE QRIVFE QRIE 0.2 QR: D2, N=50, T=4, t_2 0.1 Bias QRFE: D2, N=50, T=4, t_2 0.0 QRIVFE: D2, N=50, T=4, t_2 −0.1 QRIE: D2, N=50, T=4, t_2 5 10 15 20 Monte Carlo experiment Estimating and Testing a Quantile Regression Model with Interactive Effects 37 / 50 Models and Estimators Asymptotic Theory Application Conclusions 0.4 Comparing Methods 0.3 QR QRFE QRIVFE QRIE QRFE: D2, N=50, T=8, t_2 0.1 Bias 0.2 QR: D2, N=50, T=8, t_2 −0.1 0.0 QRIE: D2, N=50, T=8, t_2 5 10 15 20 Monte Carlo experiment Estimating and Testing a Quantile Regression Model with Interactive Effects 38 / 50 Models and Estimators Asymptotic Theory Application Conclusions 0.4 Comparing Methods 0.3 QR QRFE QRIVFE QRIE 0.2 QR: D2, N=100, T=4, t_2 0.1 Bias QRFE: D2, N=100, T=4, t_2 0.0 QRIVFE: D2, N=100, T=4, t −0.1 QRIE: D2, N=100, T=4, t_2 5 10 15 20 Monte Carlo experiment Estimating and Testing a Quantile Regression Model with Interactive Effects 39 / 50 Models and Estimators Asymptotic Theory Application Conclusions 0.4 Comparing Methods 0.3 QR QRFE QRIVFE QRIE QRFE: D2, N=100, T=8, 0.1 Bias 0.2 QR: D2, N=100, T=8, t_2 0.0 QRIE: D2, N=100, T=8, −0.1 QRIVFE: D2, N=100, T= 5 10 15 20 Monte Carlo experiment Estimating and Testing a Quantile Regression Model with Interactive Effects 40 / 50 Models and Estimators Asymptotic Theory Application Conclusions 0.4 Comparing Methods −0.1 0.0 0.1 Bias 0.2 0.3 QR QRFE QRIVFE QRIE 5 10 15 20 Monte Carlo experiment Estimating and Testing a Quantile Regression Model with Interactive Effects 41 / 50 Models and Estimators Asymptotic Theory Application Conclusions QR QRFE QRIVFE QRIE 0.0 0.1 0.2 RMSE 0.3 0.4 0.5 0.6 Comparing Methods 5 10 15 20 Monte Carlo experiment Estimating and Testing a Quantile Regression Model with Interactive Effects 42 / 50 Models and Estimators Asymptotic Theory Application Conclusions Bocconi University Estimating and Testing a Quantile Regression Model with Interactive Effects 43 / 50 Models and Estimators Asymptotic Theory Application Conclusions Data We employ data from the administrative records at Bocconi University. The data set includes: educational attainment (mean 25.455 ≈ B+ in US), class size, class composition, demographic and socioeconomic characteristics. Class size and class composition may be endogenous. We use instruments generated from a random assignment of students into classes. Class size instrumented with number of students generated by lotteries. Estimating and Testing a Quantile Regression Model with Interactive Effects 44 / 50 Models and Estimators Asymptotic Theory Application Conclusions Data We employ data from the administrative records at Bocconi University. The data set includes: educational attainment (mean 25.455 ≈ B+ in US), class size, class composition, demographic and socioeconomic characteristics. Class size and class composition may be endogenous. We use instruments generated from a random assignment of students into classes. Class size instrumented with number of students generated by lotteries. Estimating and Testing a Quantile Regression Model with Interactive Effects 44 / 50 Models and Estimators Asymptotic Theory Application Conclusions Data We employ data from the administrative records at Bocconi University. The data set includes: educational attainment (mean 25.455 ≈ B+ in US), class size, class composition, demographic and socioeconomic characteristics. Class size and class composition may be endogenous. We use instruments generated from a random assignment of students into classes. Class size instrumented with number of students generated by lotteries. Estimating and Testing a Quantile Regression Model with Interactive Effects 44 / 50 Models and Estimators Asymptotic Theory Application Conclusions Model specification We estimate the following model: yict = d 0ct α + x 0i β + f 0ct λi + uict d ct = w 0ct π1 + x 0i π2 + f 0ct λi + v ict , The λi ’s are associated with motivation and ability, and the fct ’s measure teaching quality. It is tempting to impose λi + fct as in Hanushek et al. (2003). However, high teaching quality may have a modest effect on performance among unmotivated students, while it may dramatically affect strong, motivated students. Estimating and Testing a Quantile Regression Model with Interactive Effects 45 / 50 Models and Estimators Asymptotic Theory Application Conclusions Model specification We estimate the following model: yict = d 0ct α + x 0i β + f 0ct λi + uict d ct = w 0ct π1 + x 0i π2 + f 0ct λi + v ict , The λi ’s are associated with motivation and ability, and the fct ’s measure teaching quality. It is tempting to impose λi + fct as in Hanushek et al. (2003). However, high teaching quality may have a modest effect on performance among unmotivated students, while it may dramatically affect strong, motivated students. Estimating and Testing a Quantile Regression Model with Interactive Effects 45 / 50 Models and Estimators Asymptotic Theory Application Conclusions Model specification We estimate the following model: yict = d 0ct α + x 0i β + f 0ct λi + uict d ct = w 0ct π1 + x 0i π2 + f 0ct λi + v ict , The λi ’s are associated with motivation and ability, and the fct ’s measure teaching quality. It is tempting to impose λi + fct as in Hanushek et al. (2003). However, high teaching quality may have a modest effect on performance among unmotivated students, while it may dramatically affect strong, motivated students. Estimating and Testing a Quantile Regression Model with Interactive Effects 45 / 50 Models and Estimators Asymptotic Theory Application Conclusions Model specification We estimate the following model: yict = d 0ct α + x 0i β + f 0ct λi + uict d ct = w 0ct π1 + x 0i π2 + f 0ct λi + v ict , The λi ’s are associated with motivation and ability, and the fct ’s measure teaching quality. It is tempting to impose λi + fct as in Hanushek et al. (2003). However, high teaching quality may have a modest effect on performance among unmotivated students, while it may dramatically affect strong, motivated students. Estimating and Testing a Quantile Regression Model with Interactive Effects 45 / 50 Models and Estimators Asymptotic Theory Application Conclusions Estimating and Testing a Quantile Regression Model with Interactive Effects 46 / 50 Models and Estimators Asymptotic Theory Application Conclusions Estimating and Testing a Quantile Regression Model with Interactive Effects 47 / 50 Models and Estimators Asymptotic Theory Application Conclusions Estimating and Testing a Quantile Regression Model with Interactive Effects 48 / 50 Models and Estimators Asymptotic Theory Application Conclusions Specification tests: p-values Estimating and Testing a Quantile Regression Model with Interactive Effects 49 / 50 Models and Estimators Asymptotic Theory Application Conclusions Brief summary and extensions 1 We propose a quantile approach for panel data models with interative effects and endogenous independent variables. 2 The approach is simple and seems to provide satisfactory large sample and finite sample performance measured in terms of quadratic loss. 3 Next stop: forecasting in panel data models. Preliminary evidence shows that QRIE models perform very well. Estimating and Testing a Quantile Regression Model with Interactive Effects 50 / 50