Introduction Setup Identification Aggregation Estimation Application Conclusion Who wins, who loses? Tools for distributional policy evaluation Maximilian Kasy Department of Economics, Harvard University Maximilian Kasy Who wins, who loses? Harvard 1 of 40 Introduction Setup I I Identification Aggregation Estimation Application Conclusion Few policy changes result in Pareto improvements Most generate WINNERS and LOSERS EXAMPLES: 1. Trade liberalization net producers vs. net consumers of goods with rising / declining prices 2. Progressive income tax reform high vs. low income earners 3. Price change of publicly provided good (health, education,...) Inframarginal, marginal, and non-consumers of the good; tax-payers 4. Migration Migrants themselves; suppliers of substitutes vs. complements to migrant labor 5. Skill biased technical change suppliers of substitutes vs. complements to technology; consumers Maximilian Kasy Who wins, who loses? Harvard 2 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion This implies... If we evaluate social welfare based on individuals’ welfare: 1. To evaluate a policy effect, we need to 1.1 define how we measure individual gains and losses, 1.2 estimate them, and 1.3 take a stance on how to aggregate them. 2. To understand political economy, we need to characterize the sets of winners and losers of a policy change. My objective: 1. tools for distributional evaluation 2. utilitarian framework, arbitrary heterogeneity, endogenous prices Maximilian Kasy Who wins, who loses? Harvard 3 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Proposed procedure 1. impute money-metric welfare effect to each individual 2. then: 2.1 report average effects given income / other covariates 2.2 construct sets of winners and losers 2.3 aggregate using welfare weights contrast with program evaluation approach: 1. effect on average 2. of observed outcome Maximilian Kasy Who wins, who loses? Harvard 4 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Contributions 1. Assumptions 1.1 endogenous prices / wages (vs. public finance) 1.2 utilitarian welfare (vs. labor, distributional decompositions) 1.3 arbitrary heterogeneity (vs. labor) 2. Objects of interest 2.1 disaggregated welfare ⇒ I I political economy allow reader to have own welfare weights 2.2 aggregated ⇒ policy evaluation as in optimal taxation 3. Formal results next slide Maximilian Kasy Who wins, who loses? Harvard 5 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Formal results 1. Identification 1.1 Main challenge: E [ẇ · l |w · l ] 1.2 More generally: E [ẋ |x ] causal effect of policy conditional on endogenous outcomes, 1.3 solution: tools from vector analysis, fluid dynamics 2. Aggregation social welfare & distributional decompositions 2.1 welfare weights ≈ derivative of influence function 2.2 welfare impact = impact on income - behavioral correction 3. Inference 3.1 local linear quantile regressions 3.2 combined with control functions 3.3 suitable weighted averages Maximilian Kasy Who wins, who loses? Harvard 6 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Literature 1. public - optimal taxation Samuelson (1947), Mirrlees (1971), Saez (2001), Chetty (2009), Hendren (2013), Saez and Stantcheva (2013) 2. labor - determinants of wage distribution Autor et al. (2008), Card (2009) 3. distributional decompositions Oaxaca (1973), DiNardo et al. (1996), Firpo et al. (2009), Rothe (2010), Chernozhukov et al. (2013) 4. sociology - class analysis Wright (2005) 5. mathematical physics - theory of differential forms Rudin (1991) 6. econometrics - various Koenker (2005), Hoderlein and Mammen (2007), Matzkin (2003), Altonji and Matzkin (2005) Maximilian Kasy Who wins, who loses? Harvard 7 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Notation I policy α ∈ R individuals i I potential outcome w α realized outcome w I partial derivatives ∂w := ∂ /∂ w with respect to policy ẇ := ∂α w α I density f cdf F quantile Q I wage w labor supply l consumption vector c taxes t covariates W Maximilian Kasy Who wins, who loses? Harvard 8 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Setup Assumption (Individual utility maximization) individuals choose c and l to solve max u (c , l ) c ,l s.t . c · p ≤ l · w − t (l · w ) + y0 . (1) v := max u I u, c, l, w vary arbitrarily across i I p, w , y0 , t depend on α ⇒ so do c, l, and v I u differentiable, increasing in c, decreasing in l, quasiconcave, does not depend on α Maximilian Kasy Who wins, who loses? Harvard 9 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Objects of interest Definition 1. Money metric utility impact of policy: ė := v̇ ∂y0 v 2. Average conditional policy effect on welfare: γ(y , W ) := E [ė|y , W , α] 3. Sets of winners and losers: W := {(y , W ) : γ(y , W ) ≥ 0} L := {(y , W ) : γ(y , W ) ≤ 0} 4. Policy effect on utilitarian welfare: SWF : v (.) → R ˙ = E [ω · γ] SWF Maximilian Kasy Harvard Who wins, who loses? 10 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Marginal policy effect on individuals Lemma ẏ = (l̇ · w + l · ẇ ) · (1 − ∂z t ) − ṫ + y˙0 , ė = l · ẇ · (1 − ∂z t ) − ṫ + y˙0 − c · ṗ. (2) Proof: Envelope theorem. 1. wage effect l · ẇ · (1 − ∂z t ), 2. effect on unearned income y˙0 , 3. mechanical effect of changing taxes −ṫ. 4. behavioral effect b := l̇ · w · (1 − ∂z t ) = l̇ · n, 5. price effect −c · ṗ. Income vs utility: ẏ − ė = l̇ · n + c · ṗ. Maximilian Kasy Harvard Who wins, who loses? 11 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Example: Introduction of EITC (cf. Rothstein, 2010) I Transfer income to poor mothers made contingent on labor income 1. mechanical effect > 0 if employed < 0 if unemployed 2. labor supply effect > 0 3. wage effect <0 for mothers and non-mothers I Evaluation based on 1. income (“labor”) 2. utility, assuming fixed wages (“public”) 3. utility, general model I I 1. mechanical + wage + labor supply 2. mechanical 3. mechanical + wage Case 3 looks worse than “labor” / “public” evaluations Maximilian Kasy Harvard Who wins, who loses? 12 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Identification of disaggregated welfare effects I Goal: identify γ(y, W) = E[ė|y, W, α] I Simplified case: no change in prices, taxes, unearned income no covariates I Then γ(y ) = E [l · ṅ|l · n, α] I Denote x = (l , w ). Need to identify g (x , α) = E [ẋ |x , α] (3) from f (x |α). I Made necessary by combination of 1. utilitarian welfare 2. heterogeneous wage response. Maximilian Kasy Harvard Who wins, who loses? 13 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Assume : 1. x = x (α, ε), x ∈ Rk 2. α ⊥ ε 3. x (., ε) differentiable Physics analogy: I x (α, ε): position of particle ε at time α I f (x |α): density of gas / fluid at time α , position x I ḟ change of density I h(x , α) = E [ẋ |x , α] · f (x |α): “flow density” Maximilian Kasy Harvard Who wins, who loses? 14 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Stirring your coffee I I If we know densities f (x |α), what do we know about flow g (x , α) = E [ẋ |x , α]? Problem: Stirring your coffee I does not change its density, I yet moves it around. I ⇒ different flows g (x , α) consistent with a constant density f (x |α) Maximilian Kasy Harvard Who wins, who loses? 15 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Will show: I Knowledge of f (x |α) I I I I Add to h I I I I identifies ∇ · h = ∑kj=1 ∂x j hj where h = E [ẋ |x , α] · f (x |α), identifies nothing else. h̃ such that ∇ · h̃ ≡ 0 ⇒ f (x |α) does not change “stirring your coffee” Additional conditions I I I e.g.: “wage response unrelated to initial labor supply” ⇒ just-identification of g (x , α) = E [ẋ |x , α] g j (x , α) = ∂α Q (v j |v 1 , . . . , v j −1 , .α) Maximilian Kasy Harvard Who wins, who loses? 16 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Density and flow Recall h(x , α) := E [ẋ |x , α] · f (x |α) k ∇ · h := ∑ ∂x j hj j =1 ḟ := ∂α f (x |α) Theorem ḟ = −∇ · h (4) h2+dx2h2 h1+dx1h1 h1 h2 Maximilian Kasy Harvard Who wins, who loses? 17 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Proof: 1. For some a(x ), let Z A(α) := E [a(x (α, ε))|α] = a(x (α, ε))dP (ε) Z = a(x )f (x |α)dx . 2. By partial integration: k Ȧ(α) = E [∂x a · ẋ |α] = Z ∑ ∂x j a · hj dx j =1 =− Z k a· ∑ ∂x j hj dx = − Z a · (∇ · h)dx . j =1 3. Alternatively: Z Ȧ(α) = a(x )ḟ (x |α)dx . 4. 2 and 3 hold for any a ⇒ ḟ = −∇ · h. Maximilian Kasy Harvard Who wins, who loses? 18 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion The identified set Theorem The identified set for h is given by h0 + H (5) where H = {h̃ : ∇ · h̃ ≡ 0} h (x , α) = f (x |α) · ∂α Q (v j |v 1 , . . . , v j −1 , α) 0j v j = F (x j |x 1 , . . . , x j −1 , α) Maximilian Kasy Harvard Who wins, who loses? 19 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Proof of sharpness of identified set: 1. For any h s.t. ḟ = −∇ · h construct DGP as follows 2. Let ε = x (0, ε), f (ε) = f (x |α = 0) 3. Let x (., ε) be the solution of the ODE ẋ = g (x , α), x (0, ε) = ε. (existence: Peano’s theorem) 4. ⇒ consistent with h and with f Maximilian Kasy Harvard Who wins, who loses? 20 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Theorem 1. Suppose k = 1. Then H = {h̃ ≡ 0}. (6) H = {h̃ : h̃ = A · ∇H for some H : X → R}. (7) 2. Suppose k = 2. Then where A= 0 −1 1 0 . 3. Suppose k = 3. Then H = {h̃ : h̃ = ∇ × G}. (8) where G : X → R3 . Maximilian Kasy Harvard Who wins, who loses? 21 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Proof: Poincaré’s Lemma. Figure : Incompressible flow and rotated gradient of potential h̃ = ∇ · (x 1 · exp(−x 21 − x 22 )) 2 H̃ = x 1 · exp(−x 21 − x 22 ) 1.5 1 0.5 x2 0.5 0 0 −0.5 −1 −0.5 2 1 −1.5 2 1 0 −2 −2 0 −1 −1.5 −1 −0.5 0 x 1 0.5 1 1.5 2 x2 −1 −2 −2 x1 Maximilian Kasy Harvard Who wins, who loses? 22 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Point identification Theorem Assume ∂ E [ẋ i |x , α] = 0 for j > i . ∂ xj (9) Then h is point identified, and equal to h0 as defined before. In particular g j (x , α) = E [ẋ j |x , α] = ∂α Q (v j |v 1 , . . . , v j −1 , α). Maximilian Kasy Harvard Who wins, who loses? 23 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Controls; back to γ Proposition I Suppose α ⊥ ε|W, and ∂∂x j E [ẋ i |x , W , α] = 0 for j > i. Then E [ẋ j |x , W , α] = ∂α Q (v j |v 1 , . . . , v j −1 , W , α), where v j = F (x j |x 1 , . . . , x j −1 , W , α). I If x j = n, γ(y, W) =E [l · ṅ|y , W , α] = E[l · ∂α Q(v j |v 1 , . . . , v j −1 , W , α)|y, W, α]. (10) Instrumental variables: similar (see paper) Maximilian Kasy Harvard Who wins, who loses? 24 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Problems in practice 1. Equilibrium effects: identification of causal effects using individual-level variation requires “stable unit treatment values” (SUTVA) I I Possible solution: city level variation (cf. Card, 2009) or state level variation ⇒ appropriate definition of “treatment” no issue for evaluation of historical changes 2. Unemployment: our expression for ė requires: only monetary constraints are affected by the policy change policies shifting degree of involuntary unemployment ⇒ underestimate welfare impact Maximilian Kasy Harvard Who wins, who loses? 25 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Aggregation I Relationship social welfare ⇔ distributional decompositions? I public finance welfare weights ≈ derivative of dist decomp influence functions I ˙ Alternative representations of SWF ˙ : ⇒ alternative ways to estimate SWF 1. weighted average of individual welfare effects ė, γ 2. distributional decomposition for counterfactual income ỹ (holding labor supply constant) 3. distributional decomposition of realized income minus behavioral correction Maximilian Kasy Harvard Who wins, who loses? 26 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Welfare weights and influence functions Consider θ : P y → R Theorem 1. Welfare weights: ˙ = E [ω SWF · ė] SWF (11) θ̇ = E [ω θ · ẏ ]. α α 2. Influence function: Let s(y ) = ∂α f y (y )/f y (y ). θ̇ = E [IF (y ) · s(y )] . 3. Relating the two: ω θ = ∂y IF (y ). Maximilian Kasy Harvard Who wins, who loses? 27 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Alternative representations Theorem I Assume ω SWF = ω θ = ω and ṗ = 0. I Let ỹ α = l 0 · w α − t α (l 0 · w α ) + y0α , b = l̇ · n Then ė = ỹ˙ = ẏ − b and 1. Counterfactual income distribution: ˙ = E [ω · ỹ˙ ]= E[ω · γ] SWF (12) = E [IF (y ) · s̃(y )] α = ∂α θ P ỹ . 2. Behavioral correction of distributional decomposition: ˙ = E [ω · b]. θ̇ − SWF (13) Maximilian Kasy Harvard Who wins, who loses? 28 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Estimation 1. First estimate the disaggregated welfare impact γ(y , W ) = E [ė|y , W , α] = E [l · ẇ · (1 − ∂z t ) − ṫ + y˙0 − c · ṗ|y , W , α] (14) 2. Then estimate other objects by plugging in γb: c= {(y , W ) : γb(y , W ) ≥ 0} W c= {(y , W ) : γb(y , W ) ≤ 0} L [ ˙ = EN [ωi · γb(yi , Wi )]. SWF (15) Maximilian Kasy Harvard Who wins, who loses? 29 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Estimation of g and γ 1) b v j : estimate of F (x j |x 1 , . . . , x j −1 , W , α) 1. estimated conditional quantile of x j given (W , α, b v 1, . . . , b v j −1 ) 2. estimate by local average j 3. local weights: Ki for observation i around (W , α, b v 1, . . . , b v j −1 ) 4. j b vj = j EN [Ki · 1(xi ≤ x j )] j EN [Ki ] (16) Maximilian Kasy Harvard Who wins, who loses? 30 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion 2) b g j : estimate of E [ẋ j |x , W α] 1. identified by slope of quantile regression 2. estimate by local linear Qreg j j 3. regression residual: Ui = xi − x j − g · αi j j j 4. loss function: Li = Ui · (b v j − 1(Ui ≤ 0)) 5. h i j j b g j = argmin EN Ki · Li , g 3) γb(y , W ): estimate of E [l · ṅ|y , W , α] 1. n = x j ⇒ ė = l · ṅ = l · ẋ j 2. estimate γ by weighted average γb(y , W ) = EN Ki · l · b gj EN [Ki ] Maximilian Kasy Harvard Who wins, who loses? 31 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Inference I b g j depends on data in 3 ways: 1. through x j , α , 2. quantile b vj, v j −1 ). 3. controls (b v 1, . . . , b I 1 standard, 2 negligible, 3 nasty I to avoid dealing with 3: non-analytic methods of inference I I I I bootstrap Bayesian bootstrap subsampling validity? Maximilian Kasy Harvard Who wins, who loses? 32 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Application: distributional impact of EITC I Work in progress I Meyer and Rosenbaum (2001), Rothstein (2010), Leigh (2010) I March CPS sample from IPUMS I following Leigh (2010): Variation in state top-ups of EITC across time and states I α = maximum EITC benefit available (weighted average across family sizes) Maximilian Kasy Harvard Who wins, who loses? 33 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Mapping the model to the data I Need: l , w , n, y0 , t , p and covariates I tentative mapping: 1. l: weeks worked × usual hours worked per week 2. w: income from wages + self employment + farm work /l 3. n: w × (1− marginal federal tax rate) 4. p: CPI (just normalize everything) 5. y0 : all other income 6. t: ? 7. Covariates: age, sex, race, ...? Maximilian Kasy Harvard Who wins, who loses? 34 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Conclusion and Outlook I Most policies generate winners and losers I Motivates interest in 1. disaggregated welfare effects 2. sets of winners and losers (political economy!) 3. weighted average welfare effects (optimal policy!) I Consider framework which allows for 1. endogenous prices / wages (vs. public finance) 2. utilitarian welfare (vs. labor, distributional decompositions) 3. arbitrary heterogeneity (vs. labor) Maximilian Kasy Harvard Who wins, who loses? 35 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Main results 1. Identification 1.1 Main challenge: E [ẇ · l |w · l ] 1.2 More generally: E [ẋ |x ] causal effect of policy conditional on endogenous outcomes, 1.3 solution: tools from vector analysis, fluid dynamics 2. Aggregation social welfare & distributional decompositions 2.1 welfare weights ≈ derivative of influence function 2.2 welfare impact = impact on income - behavioral correction 3. Inference 3.1 local linear quantile regressions 3.2 combined with control functions 3.3 suitable weighted averages Maximilian Kasy Harvard Who wins, who loses? 36 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Generalization of identification to dim(α) > 1 I g (x , α) = E [∂α x |x , α] ∈ Rl , h(x , α) = g (x , α) · f (x |α) I ∇ · h := (∇ · h1 , . . . , ∇ · hl ) I Most results immediately generalize I In particular Theorem ∂α f = −∇ · h (17) Maximilian Kasy Harvard Who wins, who loses? 37 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Theorem The identified set for h is contained in h0 + H (18) where H = {h̃ : ∇ · h̃ ≡ 0} h (x , α) = f (x |α) · ∂α Q (v j |v 1 , . . . , v j −1 , α) 0j v j = F (x j |x 1 , . . . , x j −1 , α) I open question: is this sharp? I does the model restrict the set of admissible g? Maximilian Kasy Harvard Who wins, who loses? 38 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion A partial answer Lemma The system of PDEs ∂α x (α) = g (α, x ) x (0) = x 0 has a local solution iff ∂α g j + ∂x g j · g (19) is symmetric ∀ j. This solution is furthermore unique. Proof: if: differentiation. only if: Frobenius’ theorem. I cf. proof of sharpness in 1-d case I Q: what is the convex hull of all such g? Maximilian Kasy Harvard Who wins, who loses? 39 of 40 Introduction Setup Identification Aggregation Estimation Application Conclusion Thanks for your time! Maximilian Kasy Harvard Who wins, who loses? 40 of 40