Nonparametric Estimators of Dose-Response Functions Nonparametric Estimators of Dose-Response Functions Michela Bia, Alessandra Mattei, Alfonso Flores-Lagunes, Carlos Flores CEPS/INSTEAD, Luxembourg Department of Statistics "G. Parenti", University of Florence State University of New York - Binghamton University of Miami IRVAPP June, 7th 2012 Nonparametric Estimators of Dose-Response Functions Motivation The evaluation process in the field of economics, sociology, law and many other settings relies on the implementation of non-experimental techniques. In many empirical studies, the units under analysis may receive different treatment levels, in which case the estimation of Dose-response Function is needed (Flores et al., 2012; Bia and Mattei, 2012; Hirano and Imbens, 2004). A key assumption (uncounfoundedness) is made, in order to adjust for systematic differences between groups receiving different levels of the treatment in a set of pre-treatment variables. Nonparametric Estimators of Dose-Response Functions Motivation The evaluation process in the field of economics, sociology, law and many other settings relies on the implementation of non-experimental techniques. In many empirical studies, the units under analysis may receive different treatment levels, in which case the estimation of Dose-response Function is needed (Flores et al., 2012; Bia and Mattei, 2012; Hirano and Imbens, 2004). A key assumption (uncounfoundedness) is made, in order to adjust for systematic differences between groups receiving different levels of the treatment in a set of pre-treatment variables. Nonparametric Estimators of Dose-Response Functions Motivation The evaluation process in the field of economics, sociology, law and many other settings relies on the implementation of non-experimental techniques. In many empirical studies, the units under analysis may receive different treatment levels, in which case the estimation of Dose-response Function is needed (Flores et al., 2012; Bia and Mattei, 2012; Hirano and Imbens, 2004). A key assumption (uncounfoundedness) is made, in order to adjust for systematic differences between groups receiving different levels of the treatment in a set of pre-treatment variables. Nonparametric Estimators of Dose-Response Functions Motivation The evaluation process in the field of economics, sociology, law and many other settings relies on the implementation of non-experimental techniques. In many empirical studies, the units under analysis may receive different treatment levels, in which case the estimation of Dose-response Function is needed (Flores et al., 2012; Bia and Mattei, 2012; Hirano and Imbens, 2004). A key assumption (uncounfoundedness) is made, in order to adjust for systematic differences between groups receiving different levels of the treatment in a set of pre-treatment variables. Nonparametric Estimators of Dose-Response Functions Reference Literature Propensity score-based methods have been widely employed in the last decades for the evaluation of causal effects, but usually confined to the binary treatment case (Rosenbaum and Rubin, 1983). Over the last years, several extensions of propensity score techniques have been proposed and developed to deal with arbitrary treatment regimes: I Imbens (2000), Lechner (2001) in the case of categorical treatment variables; I Joffe and Rosenbaum (1999) for ordinal treatment variables; I Hirano and Imbens (2004), Imai and Van Dyk (2004) and Flores et al. (2012) for continuous treatment variables. Nonparametric Estimators of Dose-Response Functions Reference Literature Propensity score-based methods have been widely employed in the last decades for the evaluation of causal effects, but usually confined to the binary treatment case (Rosenbaum and Rubin, 1983). Over the last years, several extensions of propensity score techniques have been proposed and developed to deal with arbitrary treatment regimes: I Imbens (2000), Lechner (2001) in the case of categorical treatment variables; I Joffe and Rosenbaum (1999) for ordinal treatment variables; I Hirano and Imbens (2004), Imai and Van Dyk (2004) and Flores et al. (2012) for continuous treatment variables. Nonparametric Estimators of Dose-Response Functions Reference Literature Propensity score-based methods have been widely employed in the last decades for the evaluation of causal effects, but usually confined to the binary treatment case (Rosenbaum and Rubin, 1983). Over the last years, several extensions of propensity score techniques have been proposed and developed to deal with arbitrary treatment regimes: I Imbens (2000), Lechner (2001) in the case of categorical treatment variables; I Joffe and Rosenbaum (1999) for ordinal treatment variables; I Hirano and Imbens (2004), Imai and Van Dyk (2004) and Flores et al. (2012) for continuous treatment variables. Nonparametric Estimators of Dose-Response Functions Reference Literature Propensity score-based methods have been widely employed in the last decades for the evaluation of causal effects, but usually confined to the binary treatment case (Rosenbaum and Rubin, 1983). Over the last years, several extensions of propensity score techniques have been proposed and developed to deal with arbitrary treatment regimes: I Imbens (2000), Lechner (2001) in the case of categorical treatment variables; I Joffe and Rosenbaum (1999) for ordinal treatment variables; I Hirano and Imbens (2004), Imai and Van Dyk (2004) and Flores et al. (2012) for continuous treatment variables. Nonparametric Estimators of Dose-Response Functions Reference Literature Propensity score-based methods have been widely employed in the last decades for the evaluation of causal effects, but usually confined to the binary treatment case (Rosenbaum and Rubin, 1983). Over the last years, several extensions of propensity score techniques have been proposed and developed to deal with arbitrary treatment regimes: I Imbens (2000), Lechner (2001) in the case of categorical treatment variables; I Joffe and Rosenbaum (1999) for ordinal treatment variables; I Hirano and Imbens (2004), Imai and Van Dyk (2004) and Flores et al. (2012) for continuous treatment variables. Nonparametric Estimators of Dose-Response Functions Reference Literature Propensity score-based methods have been widely employed in the last decades for the evaluation of causal effects, but usually confined to the binary treatment case (Rosenbaum and Rubin, 1983). Over the last years, several extensions of propensity score techniques have been proposed and developed to deal with arbitrary treatment regimes: I Imbens (2000), Lechner (2001) in the case of categorical treatment variables; I Joffe and Rosenbaum (1999) for ordinal treatment variables; I Hirano and Imbens (2004), Imai and Van Dyk (2004) and Flores et al. (2012) for continuous treatment variables. Nonparametric Estimators of Dose-Response Functions Our contribution to the existing literature I Hirano and Imbens (2004) address estimation of DRF by introducing the concept of the Generalized Propensity Score (GPS) and implement their approach in a parametric setting. I Flores at al. (2012) use the GPS within a weighting and introduce a Inverse Weighted (IW) kernel estimator for semiparametrically estimating the DRF. We propose two new semiparametric estimators of DRF based on spline methods, developing a set of Stata programs to I I I I I i) estimate the GPS by using generalized linear models ii) common support condition iii) test the balancing property of the estimated generalized propensity score iv) estimate the DRF using either penalized spline estimators or IW kernel estimator. Nonparametric Estimators of Dose-Response Functions Our contribution to the existing literature I Hirano and Imbens (2004) address estimation of DRF by introducing the concept of the Generalized Propensity Score (GPS) and implement their approach in a parametric setting. I Flores at al. (2012) use the GPS within a weighting and introduce a Inverse Weighted (IW) kernel estimator for semiparametrically estimating the DRF. We propose two new semiparametric estimators of DRF based on spline methods, developing a set of Stata programs to I I I I I i) estimate the GPS by using generalized linear models ii) common support condition iii) test the balancing property of the estimated generalized propensity score iv) estimate the DRF using either penalized spline estimators or IW kernel estimator. Nonparametric Estimators of Dose-Response Functions Our contribution to the existing literature I Hirano and Imbens (2004) address estimation of DRF by introducing the concept of the Generalized Propensity Score (GPS) and implement their approach in a parametric setting. I Flores at al. (2012) use the GPS within a weighting and introduce a Inverse Weighted (IW) kernel estimator for semiparametrically estimating the DRF. We propose two new semiparametric estimators of DRF based on spline methods, developing a set of Stata programs to I I I I I i) estimate the GPS by using generalized linear models ii) common support condition iii) test the balancing property of the estimated generalized propensity score iv) estimate the DRF using either penalized spline estimators or IW kernel estimator. Nonparametric Estimators of Dose-Response Functions Our contribution to the existing literature I Hirano and Imbens (2004) address estimation of DRF by introducing the concept of the Generalized Propensity Score (GPS) and implement their approach in a parametric setting. I Flores at al. (2012) use the GPS within a weighting and introduce a Inverse Weighted (IW) kernel estimator for semiparametrically estimating the DRF. We propose two new semiparametric estimators of DRF based on spline methods, developing a set of Stata programs to I I I I I i) estimate the GPS by using generalized linear models ii) common support condition iii) test the balancing property of the estimated generalized propensity score iv) estimate the DRF using either penalized spline estimators or IW kernel estimator. Nonparametric Estimators of Dose-Response Functions Our contribution to the existing literature I Hirano and Imbens (2004) address estimation of DRF by introducing the concept of the Generalized Propensity Score (GPS) and implement their approach in a parametric setting. I Flores at al. (2012) use the GPS within a weighting and introduce a Inverse Weighted (IW) kernel estimator for semiparametrically estimating the DRF. We propose two new semiparametric estimators of DRF based on spline methods, developing a set of Stata programs to I I I I I i) estimate the GPS by using generalized linear models ii) common support condition iii) test the balancing property of the estimated generalized propensity score iv) estimate the DRF using either penalized spline estimators or IW kernel estimator. Nonparametric Estimators of Dose-Response Functions Our contribution to the existing literature I Hirano and Imbens (2004) address estimation of DRF by introducing the concept of the Generalized Propensity Score (GPS) and implement their approach in a parametric setting. I Flores at al. (2012) use the GPS within a weighting and introduce a Inverse Weighted (IW) kernel estimator for semiparametrically estimating the DRF. We propose two new semiparametric estimators of DRF based on spline methods, developing a set of Stata programs to I I I I I i) estimate the GPS by using generalized linear models ii) common support condition iii) test the balancing property of the estimated generalized propensity score iv) estimate the DRF using either penalized spline estimators or IW kernel estimator. Nonparametric Estimators of Dose-Response Functions Our contribution to the existing literature I Hirano and Imbens (2004) address estimation of DRF by introducing the concept of the Generalized Propensity Score (GPS) and implement their approach in a parametric setting. I Flores at al. (2012) use the GPS within a weighting and introduce a Inverse Weighted (IW) kernel estimator for semiparametrically estimating the DRF. We propose two new semiparametric estimators of DRF based on spline methods, developing a set of Stata programs to I I I I I i) estimate the GPS by using generalized linear models ii) common support condition iii) test the balancing property of the estimated generalized propensity score iv) estimate the DRF using either penalized spline estimators or IW kernel estimator. Nonparametric Estimators of Dose-Response Functions Application (1/2) We use a data set collected by Imbens, Rubin and Sacerdote (2001). The winners of the Megabucks lottery in Massachusetts in the mid-1980’s represent the reference population. We implement our programs to semi-parametrically estimate the average potential post-winning labor earnings for each amount of the lottery prize (DRF). Nonparametric Estimators of Dose-Response Functions Application (1/2) We use a data set collected by Imbens, Rubin and Sacerdote (2001). The winners of the Megabucks lottery in Massachusetts in the mid-1980’s represent the reference population. We implement our programs to semi-parametrically estimate the average potential post-winning labor earnings for each amount of the lottery prize (DRF). Nonparametric Estimators of Dose-Response Functions Application (1/2) We use a data set collected by Imbens, Rubin and Sacerdote (2001). The winners of the Megabucks lottery in Massachusetts in the mid-1980’s represent the reference population. We implement our programs to semi-parametrically estimate the average potential post-winning labor earnings for each amount of the lottery prize (DRF). Nonparametric Estimators of Dose-Response Functions Application (2/2) The assignment of the prize is random, but unit and item nonresponse led to a selected sample where the amount of the prize received is no longer independent of background characteristics. As in the binary treatment case, the extent to which the bias generated by unobservable confounding factors is reduced heavily depends on the richness of the pre-treatment variables available in the empirical study. Nonparametric Estimators of Dose-Response Functions Application (2/2) The assignment of the prize is random, but unit and item nonresponse led to a selected sample where the amount of the prize received is no longer independent of background characteristics. As in the binary treatment case, the extent to which the bias generated by unobservable confounding factors is reduced heavily depends on the richness of the pre-treatment variables available in the empirical study. Nonparametric Estimators of Dose-Response Functions Estimation Strategy [ outline ] Estimation Strategy The Stata program Illustrative Example On the agenda Nonparametric Estimators of Dose-Response Functions Estimation Strategy Basic Framework (1/2) We refer to the potential outcome approach to causal inference (Rubin, 1974, 1978). We estimate a continuous DRF that relates each value of the dose, i.e., lottery prize level, to the post-winning labor earnings. Formally, consider a set of N individuals, and denote each of them by subscript i: i = 1, . . . , N. For each individual i, we observe a vector of pre-treatment covariates, Xi , the received lottery prize level, Ti , and the correspondent value of the outcome for this treatment level, Yi = Yi (Ti ). Nonparametric Estimators of Dose-Response Functions Estimation Strategy Basic Framework (1/2) We refer to the potential outcome approach to causal inference (Rubin, 1974, 1978). We estimate a continuous DRF that relates each value of the dose, i.e., lottery prize level, to the post-winning labor earnings. Formally, consider a set of N individuals, and denote each of them by subscript i: i = 1, . . . , N. For each individual i, we observe a vector of pre-treatment covariates, Xi , the received lottery prize level, Ti , and the correspondent value of the outcome for this treatment level, Yi = Yi (Ti ). Nonparametric Estimators of Dose-Response Functions Estimation Strategy Basic Framework (1/2) We refer to the potential outcome approach to causal inference (Rubin, 1974, 1978). We estimate a continuous DRF that relates each value of the dose, i.e., lottery prize level, to the post-winning labor earnings. Formally, consider a set of N individuals, and denote each of them by subscript i: i = 1, . . . , N. For each individual i, we observe a vector of pre-treatment covariates, Xi , the received lottery prize level, Ti , and the correspondent value of the outcome for this treatment level, Yi = Yi (Ti ). Nonparametric Estimators of Dose-Response Functions Estimation Strategy Basic Framework (1/2) We refer to the potential outcome approach to causal inference (Rubin, 1974, 1978). We estimate a continuous DRF that relates each value of the dose, i.e., lottery prize level, to the post-winning labor earnings. Formally, consider a set of N individuals, and denote each of them by subscript i: i = 1, . . . , N. For each individual i, we observe a vector of pre-treatment covariates, Xi , the received lottery prize level, Ti , and the correspondent value of the outcome for this treatment level, Yi = Yi (Ti ). Nonparametric Estimators of Dose-Response Functions Estimation Strategy Basic Framework 2/2 Average dose-response function µ(t) = E [Yi (t)] Unconfoundedness assumption Yi (t) ⊥ Ti |Xi for all t ∈ T Generalized Propensity Score r (T , X ) = fT |X (T |x) The GPS is like a balancing score (e.g., Rosenbaum and Rubin, 1983)...within strata with the same value of r (t, X ), the probability that T = t does not depend on the value of X . Nonparametric Estimators of Dose-Response Functions Estimation Strategy Basic Framework 2/2 Average dose-response function µ(t) = E [Yi (t)] Unconfoundedness assumption Yi (t) ⊥ Ti |Xi for all t ∈ T Generalized Propensity Score r (T , X ) = fT |X (T |x) The GPS is like a balancing score (e.g., Rosenbaum and Rubin, 1983)...within strata with the same value of r (t, X ), the probability that T = t does not depend on the value of X . Nonparametric Estimators of Dose-Response Functions Estimation Strategy Basic Framework 2/2 Average dose-response function µ(t) = E [Yi (t)] Unconfoundedness assumption Yi (t) ⊥ Ti |Xi for all t ∈ T Generalized Propensity Score r (T , X ) = fT |X (T |x) The GPS is like a balancing score (e.g., Rosenbaum and Rubin, 1983)...within strata with the same value of r (t, X ), the probability that T = t does not depend on the value of X . Nonparametric Estimators of Dose-Response Functions Estimation Strategy Basic Framework 2/2 Average dose-response function µ(t) = E [Yi (t)] Unconfoundedness assumption Yi (t) ⊥ Ti |Xi for all t ∈ T Generalized Propensity Score r (T , X ) = fT |X (T |x) The GPS is like a balancing score (e.g., Rosenbaum and Rubin, 1983)...within strata with the same value of r (t, X ), the probability that T = t does not depend on the value of X . Nonparametric Estimators of Dose-Response Functions Estimation Strategy Basic Framework 2/2 Average dose-response function µ(t) = E [Yi (t)] Unconfoundedness assumption Yi (t) ⊥ Ti |Xi for all t ∈ T Generalized Propensity Score r (T , X ) = fT |X (T |x) The GPS is like a balancing score (e.g., Rosenbaum and Rubin, 1983)...within strata with the same value of r (t, X ), the probability that T = t does not depend on the value of X . Nonparametric Estimators of Dose-Response Functions Estimation Strategy Treatment Effect Estimation: Example Normal distribution of the treatment given the covariates f (T )|X ∼ N b0 + b1 X , σ 2 Estimated Generalized Propensity Score R̂i = √ 1 2πσ̂ 2 exp(− 1 (f (Ti ) − b̂0 − b̂1 Xi )2 ) 2σ̂ 2 Nonparametric Estimators of Dose-Response Functions Estimation Strategy Treatment Effect Estimation: Example Normal distribution of the treatment given the covariates f (T )|X ∼ N b0 + b1 X , σ 2 Estimated Generalized Propensity Score R̂i = √ 1 2πσ̂ 2 exp(− 1 (f (Ti ) − b̂0 − b̂1 Xi )2 ) 2σ̂ 2 Nonparametric Estimators of Dose-Response Functions Estimation Strategy Parametric Approach Estimation through a quadratic approximation: E (Yi |Ti , R̂i ) = α0 + α1 Ti + α2 Ti2 + α3 R̂i + α4 Rˆi2 + α5 Ti R̂i Average potential outcome at treatment level t, given αˆ0 ,, αˆ2 ...αˆ5 P E\ [Y (t)] = N1 ni=1 (αˆ0 + αˆ1 t + αˆ2 t 2 + αˆ3 r̂ (t, Xi ) + ... + αˆ5 tr̂ (t, Xi )) Nonparametric Estimators of Dose-Response Functions Estimation Strategy Parametric Approach Estimation through a quadratic approximation: E (Yi |Ti , R̂i ) = α0 + α1 Ti + α2 Ti2 + α3 R̂i + α4 Rˆi2 + α5 Ti R̂i Average potential outcome at treatment level t, given αˆ0 ,, αˆ2 ...αˆ5 P E\ [Y (t)] = N1 ni=1 (αˆ0 + αˆ1 t + αˆ2 t 2 + αˆ3 r̂ (t, Xi ) + ... + αˆ5 tr̂ (t, Xi )) Nonparametric Estimators of Dose-Response Functions Estimation Strategy Flexible specification of the Generalized Propensity Score We estimate the conditional distribution of the treatment given the covariates under 4 different distributional assumptions of the treatment conditional on the covariates: I i) Gaussian Distribution, I ii) Inverse Gaussian Distribution I iii) Gamma Distribution I iv) Beta Distribution Nonparametric Estimators of Dose-Response Functions Estimation Strategy Flexible specification of the Generalized Propensity Score We estimate the conditional distribution of the treatment given the covariates under 4 different distributional assumptions of the treatment conditional on the covariates: I i) Gaussian Distribution, I ii) Inverse Gaussian Distribution I iii) Gamma Distribution I iv) Beta Distribution Nonparametric Estimators of Dose-Response Functions Estimation Strategy Flexible specification of the Generalized Propensity Score We estimate the conditional distribution of the treatment given the covariates under 4 different distributional assumptions of the treatment conditional on the covariates: I i) Gaussian Distribution, I ii) Inverse Gaussian Distribution I iii) Gamma Distribution I iv) Beta Distribution Nonparametric Estimators of Dose-Response Functions Estimation Strategy Flexible specification of the Generalized Propensity Score We estimate the conditional distribution of the treatment given the covariates under 4 different distributional assumptions of the treatment conditional on the covariates: I i) Gaussian Distribution, I ii) Inverse Gaussian Distribution I iii) Gamma Distribution I iv) Beta Distribution Nonparametric Estimators of Dose-Response Functions Estimation Strategy Flexible specification of the Generalized Propensity Score We estimate the conditional distribution of the treatment given the covariates under 4 different distributional assumptions of the treatment conditional on the covariates: I i) Gaussian Distribution, I ii) Inverse Gaussian Distribution I iii) Gamma Distribution I iv) Beta Distribution Nonparametric Estimators of Dose-Response Functions Estimation Strategy Nonparametric Approach: Spline Estimators (1/4) Following Ruppert et al (2003), we use bivariate basis function. The additive model for truncated lines is: t bi + E (Yi |Ti , R̂i ) = a0 +at Ti +ar R K X k=1 r ukt (Ti −kkt )+ + K X bi −k r )+ ukr (R k k=1 where k1t < . . . < kKt t and k1 < . . . < kKr r are K t and K r distinct bi . knots in the support of T and the support ofthe estimated GPS, R Nonparametric Estimators of Dose-Response Functions Estimation Strategy Nonparametric Approach: Spline Estimators (2/4) We also allowed for interactions in the model by adding the following product terms: bi + λTi R bi + E (Yi |Ti , R̂i ) = a0 + at Ti + ar R t K X r ukt (Ti − kkt )+ k=1 + K X bi − k r )+ ukr (R k k=1 t r + K X bi (Ti vkr R k=1 Kt X Kr X k=1 − kkt )+ + K X b i − k r )+ + vkt Ti (R k k=1 tr t r b vkk 0 (Ti − kk )+ (Ri − kk 0 )+ k 0 =1 The model is obtained from the basis functions 1, Ti t, (Ti − k1t ) . . . , (Ti − kKt t ) b t , (R bi − k r ) . . . , (R bi − k t r ) 1, R 1 i K (1) Nonparametric Estimators of Dose-Response Functions Estimation Strategy Nonparametric Approach: Spline Estimators (3/4) An alternative to tensor product splines is given by the so-called radial basis functions, which are basis functions of the form C (k(t, r )0 − (k, k 0 )0 k) for some univariate function C . Here we consider the following function C t ! t 2 t t t t k k k − = − log − r r kr r kr kr where k · k is the Euclidean norm. Nonparametric Estimators of Dose-Response Functions Estimation Strategy Nonparametric Approach: Spline Estimators (4/4) We assume that K X bi +λTi R̂i + E (Yi |Ti , R̂i ) = a0 +at Ti +ar R uk C k=1 t ! Ti kk − R̂i kkr (2) where u1 , · · · , uk are random variables with mean 0, and − 12 0 − 12 2 )(Ω variance-covariance matrix Cov (u) = σ (Ω u k k ) , with " t t !# k kk 0 k − . Ωk = C kr kr0 k bt R i k 1≤k,k 0 ≤K Let =b r (t, Xi ) denote the estimated score at a specific treatment level, t. Given the estimated parameters of the additive, tensor or radial regression functions, the average potential outcome bt. at treatment level t is estimated averaging as over R i Nonparametric Estimators of Dose-Response Functions Estimation Strategy Nonparametric approach: Inverse-Weighting Average potential outcome estimated by I-W Kernel Estimator D0 (t)S2 (t) − D1 (t)S1 (t) E\ [Y (t)] = S0 (t)S2 (t) − S12 (t) Sj (t) = N X k̃hX (Ti − t)(Ti − t)j i=1 Dj (t) = N X k̃hX (Ti − t)(Ti − t)j Yi i=1 k̃hX (Ti − t) = Kh (Ti − t) Rˆt i Nonparametric Estimators of Dose-Response Functions Estimation Strategy Nonparametric approach: Inverse-Weighting Average potential outcome estimated by I-W Kernel Estimator D0 (t)S2 (t) − D1 (t)S1 (t) E\ [Y (t)] = S0 (t)S2 (t) − S12 (t) Sj (t) = N X k̃hX (Ti − t)(Ti − t)j i=1 Dj (t) = N X k̃hX (Ti − t)(Ti − t)j Yi i=1 k̃hX (Ti − t) = Kh (Ti − t) Rˆt i Nonparametric Estimators of Dose-Response Functions Estimation Strategy Nonparametric approach: Inverse-Weighting Average potential outcome estimated by I-W Kernel Estimator D0 (t)S2 (t) − D1 (t)S1 (t) E\ [Y (t)] = S0 (t)S2 (t) − S12 (t) Sj (t) = N X k̃hX (Ti − t)(Ti − t)j i=1 Dj (t) = N X k̃hX (Ti − t)(Ti − t)j Yi i=1 k̃hX (Ti − t) = Kh (Ti − t) Rˆt i Nonparametric Estimators of Dose-Response Functions Estimation Strategy Nonparametric approach: Inverse-Weighting Average potential outcome estimated by I-W Kernel Estimator D0 (t)S2 (t) − D1 (t)S1 (t) E\ [Y (t)] = S0 (t)S2 (t) − S12 (t) Sj (t) = N X k̃hX (Ti − t)(Ti − t)j i=1 Dj (t) = N X k̃hX (Ti − t)(Ti − t)j Yi i=1 k̃hX (Ti − t) = Kh (Ti − t) Rˆt i Nonparametric Estimators of Dose-Response Functions The Stata program [ outline ] Estimation Strategy The Stata program Illustrative Example On the agenda Nonparametric Estimators of Dose-Response Functions The Stata program The syntax DRF varlist , outcome(varname) treatment(varname) gpscore(newvar) cutpoints(varname) index(string) nq_gps(num) method(type) family(familyname) link(linkname) vce(string) common(num) test(type) tpoints(vector) detail delta(num) options for IW Kernel bandwidth() options for mtspline degree1(num) degree2(num) nknots1(num) nknots2(num) knots1(numlist) knots2(numlist) additive options for radialpspline nknots(num) knots(numlist) standardized options for both mtspline and radialpspline estopts(string) Nonparametric Estimators of Dose-Response Functions Illustrative Example [ outline ] Estimation Strategy The Stata program Illustrative Example On the agenda Nonparametric Estimators of Dose-Response Functions Illustrative Example The application on the Lottery Sample (1/2) The sample we use in this analysis is the “winners” sample of 237 individuals who won a major prize in the lottery. The outcome of interest is “year6” (earnings six years after winning the lottery), and the treatment is “prize”, the prize amount. Control variables are age, gender, years of high school, years of college, winning year, number of tickets bought, working status after the winning, and earnings s years before winning the lottery, s = 1, 2, . . . , 6. Nonparametric Estimators of Dose-Response Functions Illustrative Example The application on the Lottery Sample (2/2) DRF agew ownhs owncoll male tixbot workthen yearm1 yearm2 yearm3 yearm4, outcome(year6) treatment(prize) gpscore(gps) method(radialpspline) family(gaussian) link(log) numoverlap(3) nknots(4) tpoints(tp) flag(1) cutpoints(cut) index(p50) nq_gps(5) detail delta(1) DRF agew ownhs owncoll male tixbot workthen yearm1 yearm2 yearm3 yearm4, outcome(year6) treatment(prize) gpscore(gps) method(iwkernel) family(gaussian) link(log) numoverlap(3) tpoints(tp) flag(1) cutpoints(cut) index(p50) nq_gps(5) detail delta(1) Nonparametric Estimators of Dose-Response Functions Illustrative Example Results Dose-response function and marginal treatment effect estimation Penalized Spline Method 15 14 13 12 11 Dose−response function 16 15 10 14 13 12 11 10 9 10 20 30 40 50 60 Treatment 70 80 90 100 10 20 I−Weighting Kernel Method 30 40 50 60 Treatment 70 80 90 0 −.2 −.3 70 80 90 100 60 80 100 Radial Spline Method Derivative 0 −.1 −.1 −.2 −.3 50 60 Treatment 40 .1 −.2 40 20 Treatment 0 Derivative 0 30 11 Penalized Spline Method −.1 20 12 100 .1 10 13 9 0 .1 0 14 10 9 0 Derivative Radial Spline Method 16 15 Dose−response function Dose−response function I−Weighting Kernel Method 16 −.3 0 10 20 30 40 50 60 Treatment 70 80 90 100 0 10 20 30 40 50 60 Treatment 70 80 90 100 Nonparametric Estimators of Dose-Response Functions Illustrative Example Bootstrap bootstrap "DRF agew ownhs owncoll male tixbot workthen yearm1 yearm2 yearm3 yearm4, outcome(year6) treatment(prize) gpscore(gps) method(radialpspline) family(gaussian) link(log) numoverlap(3) nknots(4) tpoints(tp) flag(1) cutpoints(cut) index(p50) nq_gps(5) detail delta(1)" _b, reps (50) bootstrap "DRF agew ownhs owncoll male tixbot workthen yearm1 yearm2 yearm3 yearm4, outcome(year6) treatment(prize) gpscore(gps) method(iwkernel) family(gaussian) link(log) numoverlap(3) tpoints(tp) flag(1) cutpoints(cut) index(p50) nq_gps(5) detail delta(1)" _b, reps (50) Nonparametric Estimators of Dose-Response Functions Illustrative Example Bootstrap results 95% confidence intervals: Dose-response function and marginal treatment effect estimation Penalized Spline Method 30 30 30 25 20 15 10 5 Dose−response function 35 25 20 15 10 5 25 20 15 10 5 0 0 0 −5 −5 −5 −10 −10 −10 0 10 20 30 40 50 60 Treatment 70 80 90 100 0 10 20 I−Weighting Kernel Method 30 40 50 60 Treatment 70 80 90 100 0 Penalized Spline Method .6 .4 .4 .2 −.2 −.4 −.6 −.6 −.8 −.8 −.4 −.6 −.8 −1 −1 −1 −1.2 −1.2 20 30 40 50 60 Treatment 70 80 90 100 100 −.2 −1.2 10 80 0 Derivative −.4 60 .2 0 Derivative 0 −.2 40 Radial Spline Method .6 .4 0 20 Treatment .6 .2 Derivative Radial Spline Method 35 Dose−response function Dose−response function I−Weighting Kernel Method 35 0 10 20 30 40 50 60 Treatment 70 80 90 100 0 10 20 30 40 50 60 Treatment 70 80 90 100 Nonparametric Estimators of Dose-Response Functions On the agenda [ outline ] Estimation Strategy The Stata program Illustrative Example On the agenda Nonparametric Estimators of Dose-Response Functions On the agenda Possible extension I Non parametric specification of the generalized propensity score I Basic assumption (what about relaxing CIA condition for example?) Nonparametric Estimators of Dose-Response Functions On the agenda Possible extension I Non parametric specification of the generalized propensity score I Basic assumption (what about relaxing CIA condition for example?) Nonparametric Estimators of Dose-Response Functions On the agenda Possible extension I Non parametric specification of the generalized propensity score I Basic assumption (what about relaxing CIA condition for example?) Nonparametric Estimators of Dose-Response Functions On the agenda Possible extension I Non parametric specification of the generalized propensity score I Basic assumption (what about relaxing CIA condition for example?) Nonparametric Estimators of Dose-Response Functions On the agenda References Bia, M. and Mattei, A. A Stata package for the estimation of the Dose Response Function through adjustment for the generalized propensity score. The Stata Journal, 8(3), 354 - 373, (2008). Bia, M. and Mattei, A. Assessing the effect of the amount of financial aids to Piedmont firms using the generalized propensity score. Statistical Methods and Applications, forthcoming, (2012). Flores, C. A., Flores-Lagunes, A., Gonzalez, A. and Neumann, T.C. Estimating the effects of length of exposure to a training program: The case of Job Corps. The Review of Economics and Statistics, 3, (2012). Joffe, M. M. and Rosenbaum, P. R. Invited commentary: Propensity scores. American Journal of Epidemiology, 150, 327-333, (1999). Nonparametric Estimators of Dose-Response Functions On the agenda References Hirano, K. and Imbens, G. The Propensity score with continuous treatment. In Applied Bayesian Modelling and Causal Inference from Missing Data Perspectives (eds A. Gelman and X.L. Meng), Wiley,(2004). Imbens, G. W. The Role of the Propensity Score in Estimating Dose-Response Functions. Biometrika, 87, 706-710, (2000). Imai, K. and Van Dyk, D. A. Causal treatment with general treatment regimes: Generalizing the Propensity Score. Journal of the American Statistical Association, 99(467), 854-866, (2004). Lechner, M. Identification and Estimation of Causal Effects of Multiple Treatments under the Conditional Independence Assumption. In Econometric Evaluations of Active Labor Market Policies in Europe (eds M. Lechner and F. Pfeiffer), Heidelberg, Physica, (2001).
© Copyright 2025 Paperzz