Introduction The Problem Identification Data Results Conclude Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) March 27, 2013 Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Motivation I I I The ratio of boys to girls at birth in China, India, Taiwan and South Korea continuously rise even with rapid economic growth. Domestic inequality between boys and girls remains a relevant issue particularly in regions with cultural preference for boys. Social scientists have long been interested in how family composition affects human capital formation, but most of the previous studies focus on one channel – either total number of children (sibsize) or sibling sex composition – taking the other channel as absent or fixed. I I Quality and quality trade-off: Rosenzweig and Wolpin (1980), Black, Devereux and Salvanes (2005), Angrist, Lavy and Schlosser (2010) Sibling rivalry/spillover: Garg and Moduch (1998), Butcher and Case (1994), Dahl and Moretti (2008), Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Motivation I But these two channels cannot be truly separated in effect, because sibling sex composition affects sibsize if parents prefer a specific sibling-sex composition. I I I American parents of same-sex siblings tend to have an additional child (Angrist and Evans 1998). Taiwanese parents of two girls have an average of 0.53 additional children over those of two boys. In this paper we use a decomposition method to distinguish extensive from intensive margins. I I Extensive margin or indirect effect: sibling sex composition affects children’s human capital formation, by changing fertility choice. Intensive margin or direct effect: sibling sex composition affects children’s human capital formation, not by changing fertility choice. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude The Research Question I I I I We estimate the effects of family composition on formation of human capital. As starter we study only the first child with one or more siblings. We can extend the model to other parities. Human capital formation of the first child (Y ) is measured by the child’s university attainment or SAT scores. Given the firstborn’s gender, family composition is described by the gender of the next sibling (Boy2nd or B) and the number of children (or Sibsize or N ): Y = f (B, N, X) + where is the error term, and X are covariates. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude The Identification Problem If sibsize (N) and sibling gender (B) are both exogenous and B does not affect N , then an ordinary least squares (OLS) analysis would have worked. Boy2nd (B) Education (Y) Sibsize (N) Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude The Existing Literature Sibling rivalry I Sibling rivalry is conditional on sibsize. Having a son (versus a daughter) may lower parental investment in the other children if parents have resource constraints and a preference for sons (Parish-Willis 1993; Garg-Morduch 1998). Sibling feedback/spillover I Sibling feedback/spillover is conditional on sibsize. Their having a brother rather than a sister may increase parents’ investment in the sibling because of externalities. I I Gender roles and reference groups: Koch (1955), Butcher-Case (1994), Kaestner’s (1997) reanalysis of the Butcher-Case study “Intellectual environment”: Zajonc (1976) In all of the previous studies, sibling sex composition is taken as an intervention variable, and sibsize an exogenous control variable. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude The Existing Literature Conventional methods require the following assumptions: (1) The number of children is predetermined, independent of sibling sex composition. I But parents with no son are more likely to go on to have an additional child. (2) The number of children is exogenous I But sibsize and children’s human capital formation are related to unobserved parental backgrounds. (3) There is no sex selective abortion; sibling gender composition is assigned randomly. I But in regions with strong demand for sons, this assumption is too strong if ultrasound is widely available. These limitations have not been resolved in the previous literature. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude In this paper (1) We clarify parameters of interest: I I Observed sibsize cannot be fixed, but potential sibsize can be fixed conceptually. Pearl’s (2001) and VanderWeele’s (2013) conceptual models (2) We further apply instrumental-variable methods to correct for the endogenous mediator or post-intervention variable. (3) We use a unique administrative data set that covers pre-ultrasound periods and exhibits normal sex ratios. The key finding: After correcting for endogenous sibsize, we find the previous estimates are not robust. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Notations Let Y denote the observed educational outcomes and let N denote observed sibsize, N ∈ N ≡ {2, 3, ..., n̄}. We use capital letters to denote random variables, and use lower-case letters to denote their realized values. Using Rubin’s (1974) counterfactual notations, we define: I N0 , N1 : potential sibsize of the firstborn given the gender of the next child b = 0, 1. I Y0 , Y1 : potential outcome of the firstborn given the gender of the next child b = 0, 1. I Y0n , Y1n : potential outcome of the firstborn given sibsize n and sibling gender b = 0, 1. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Notations The relationship between potential and observed outcomes satisfies the “consistency” condition (Robins 1987). Assumption (Consistency) I I N = BN1 + (1 − B)N0 and Y = BY1 + (1 − B)Y0 . Y0 = Y0N0 , Y1 = Y1N1 . Notably, we never observe (Y0N1 , Y1N0 ). Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Notations The previous literature relies on the condition of randomized intervention, which we maintain in this paper. For ease of exposition, we suppress the notation of covariates exogenous X throughout the paper although our entire analysis is conditional on X. Assumption (Randomized Intervention) Conditional covariates X, we assume: (a) (Y0 , Y1 ) ⊥ B, (b) 0 < P r{B = 1} < 1. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Conventional Measures Conventional measures for the effect of the gender of the next child or the intensive margins (CIM) take observed sibsize N as an exogenous and predetermined covariate. Given N = n, define CIM (n) ≡ E[Y1 − Y0 |N = n] = E[Y1 |B = 1, N = n] − E[Y0 |B = 0, N = n] = E[Y1 |B = 1, N1 = n] − E[Y0 |B = 0, N0 = n] = E[Y |B = 1, N = n] − E[Y |B = 0, N = n] where I the second equality is because B is a randomized experiment; I the third equality requires consistency and assumes no extensive margin (N0 = N1 = n), which implicitly assumes that sibsize is exogenous and predetermined; I the fourth equation is from consistency. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Problems with the Conventional Measure The mediator variable Nb cannot be taken as a predetermined control, because its potential value depends on intervention. I Angrist and Pischke (2008) call it a “bad-control problem,” I while Heckman and Vytlacil (2007) a “feedback” issue. I Griliches and Mason (1972) I Chamberlain (1977, 1978) Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Problems with the Conventional Measure The causal-inference literature has noted the problem with the conventional methods: I Judd (1981) and Robins and Greenland (1992): In the absence of additional assumptions, direct and indirect effects are not identified even in randomized experiments. I Pearl (2001): An indirect effect, driven by a change in an intervention, is typically ill-defined in the conventional framework. I Omission of interactions between the effect of potential sibsize and the effect of sibling gender while fixing potential sibsize leads to biased results. I I VanderWeele and Vansteelandt (2009): Conventional methods that assume linear models with no interactions are biased. VanderWeele (2013) propose decomposition methods using non-parametric models. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Initial Assumptions In this paper we consider a model with a post-treatment variable (N ) and a randomized intervention (B), satisfying the following conditions, in addition to consistency and randomized intervention: Assumption (Additional Conditions for Randomized Intervention) Conditional covariates X, we assume: (a) (N0 , N1 ) ⊥ B, (b) (Y0n , Y1n ) ⊥ B|(N0 , N1 ), Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Redefining Parameters of Interest We redefine the parameters using the conditional expectation function of counter-factual outcomes, given potential sibsize: I Given sibling gender b = 0 or 1, the intensive margin (or called ”direct effect”) is conditional on the potential sibsize n. IMb (nb ) ≡ Y1nb − Y0nb . The average intensive margin is averaging over all possible values of potential sibsize n ∈ N , given the probability mass function of potential sibsize p(n|b) ≡ P r{Nb = n|b}, given b. X AIMb ≡ (Y1n − Y0n )p(n|b). n∈N Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Redefining Parameters of Interest I Given sibling gender b = 0 or 1, the extensive margin (or called ”indirect effect”) is measured by varying the potential sibsize across sibling genders. EMb (n0 , n1 ) ≡ Ybn1 − Ybn0 . The average extensive margin is X AEMb ≡ Ybn [p(n|1) − p(n|0)]. n∈N I Notably, unlike the extensive margins, the quality-quantity trade-off is measured by the effect of an increment in observed sibsize, instead of comparing potential sibsize. QQb = Yb,n+1 − Ybn , Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Decomposition I The total effect is T E(n0 , n1 ) ≡ Y1n1 − Y0n0 = (Y1n1 − Y0n1 ) + (Y0n1 − Y0n0 ) = IM1 (n1 ) + EM0 (n0 , n1 ) = IM1 (n0 ) + EM1 (n0 , n1 ) − ∆(n0 , n1 ), (1) where the adjustment term is the difference in the extensive margin across sibling genders; that is, an interaction between sibsize and sibling gender. ∆(n0 , n1 ) ≡ EM1 (n0 , n1 ) − EM0 (n0 , n1 ). I Likewise, we can write T E(n0 , n1 ) = IM0 (n0 ) + EM1 (n0 , n1 ) = IM0 (n0 ) + EM0 (n0 , n1 ) + ∆(n0 , n1 ). (2) Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Decomposition I From equations (1) and (2), the decomposition of total effect can be rewritten as follows, given b = 0 or 1: T E(n0 , n1 ) = IMb (n0 ) + EMb (n0 , n1 ) + (1 − 2b)∆(n0 , n1 ). I The average total effect and the average adjustment are X AT E ≡ E[Y1 − Y0 ] = [Y1n p(n|1) − Y0n p(n|0)] n∈N ¯ ≡ AEM1 − AEM0 = ∆ X (Y1n − Y0n )[p(n|1) − p(n|0)] n∈N Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Bias in Conventional Measure for Intensive Margin Conventional methods ignore the fact that the mediator variable Nb is affected by the intervention variable b. Its definition is given by observed sibsize and by presuming N1 = N0 = n. CIM (n) = T E(n, n) = IMb (n) + EMb (n, n) + (1 − 2b)∆(n, n) {z } | Bias Both equalities are derived by construction. I The conventional measure CIM is unbiased only if N is not affected by sibling gender, EM0 = EM1 = 0 = ∆. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Modelling Counterfactuals Given a Binary Mediator I I I For purposes of exposition, we focus on the case of a binary mediator, N = M oreT han2 = 0 or 1. The method is ready to be extended to cases of a multi- valued mediator (e.g., Sibsize). We can identify AIM and AEM if we can identify Ybn and p(n|b), both of which can be linked to the expecated values of observables, by the condition of consistency. E[Ybn |b, n] = E[Y |b, n] E[Nb |b] = E[N |b] = P r{N = 1|b} = p(n|b) ≡ n̄b . Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Parameters of interest if no interaction I We first consider a simple model of the counterfactuals Ybn , where we temporarily assume no interaction between the treatment and the mediator. Again, we suppress the notation of covariates for ease of exposition. E[Ybn |b, n] = E[Y |b, n] = β0 + β1 n + β2 b. I Greek letters are coefficients. Given this model, the parameters of interest can be spelled out: I I I I I Average intensive margin AIM0 = AIM1 = CIM = β2 . Average extensive margin AEM0 = AEM1 = β1 (n̄1 − n̄0 ). If n̄1 = n̄0 ), then AT E = β1 (n̄1 − n̄0 ) + β2 = β2 . ¯ = 0. Average adjustment term ∆ In this case CIM is an unbiased measure for AIM0 and AIM1 . Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Adding an interaction between sibsize and sibling gender Now we consider a model with an interaction term: E[Ybn |b, n] = E[Y |b, n] = β0 + β1 n + β2 b + β3 nb. The parameters of interest now become the following: I Conventional measure CIM (n) = β2 + β3 n. I Average intensive margin AIMb = β2 + β3 n̄b . I Average extensive margin AEMb = (β1 + β3 b)(n̄1 − n̄0 ). ¯ = β3 (n̄1 − n̄0 ). I Average adjustment term ∆ I Average total effect AT E = β1 (n̄1 − n̄0 ) + β2 + β3 n̄1 . I Effect of a one-unit change in the mediator on outcomes is β1 + β3 b, not necessarily equal to AEMb , unless n̄1 − n̄0 = 1. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Endogeneity of the Mediator Identification strategies When the mediator N (that is, fertility choice) is endogenous, conventional estimation methods are biased. To address this, we use conventional 2SLS methods to characterize the counterfacturals: Ybn = β0 + β1 n + β2 b + β3 nb + V Nb (z) = α0 + α1 z + α2 b + α3 zb + U where V and U are error terms. Greek letters are coefficients. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Endogeneity of the Mediator Identification strategies Assume there exists a valid instrument Z = z for the mediator variable, satisfying the following assumptions (conditional on X in the background): Assumption (Validity of the Instrument) (a) Independence: ({Ybn }n∈N , {Nb (z)}z∈Z ) ⊥ Z, given B = b; (b) Relevance: E(N |B, Z) = P (B, Z) is a non-degenerate function Z, given B. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Endogeneity of the Mediator Identification strategies I To identify AIM and AEM, we need to derive n̄1 ,n̄0 and n̄1 − n̄0 from the first-stage estimates: n̄b = α0 + α1 z̄ + α3 z̄b n̄1 − n̄0 = α2 + α3 z̄ I Let Nb (z) denote the potential fertility choice, given sibling gender b and instrument z. By a simple extension from Imbens and Angrist’s (1994) results, the average effect of sibsize on outcome can be identified, given B = b: E[Yb1 − Yb0 |Nb (0) = 0, Nb (1) = 1] E[Y |Z = 1, B = b] − E[Y |Z = 0, B = b] = E[N |Z = 1, B = b] − E[N |Z = 0, B = b] Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Examples Examples I I I One important example is the model by Butcher and Case (1994) and Kaestner (1997), who estimate the effect of sibling sex composition on children’s education. This requires an imposition that sibling gender has no effect on sibsize; that is, AEM=0. Thus CIM (n̄) = β2 + β3 n̄ = T E = AIM. Anysister Education Sibsize Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Examples Examples I I Another important example is Angrist and Evans (1998). To identify the effect of fertility choice on parents’ labor supply, they use the Samesex indicator for the first two births to instrument fertility choice. This requires an assumption that Samesex does not affect parental labor supply; that is, AIM=0. Mother's labor Supply SameSex MoreThan2 Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Examples Examples Angrist & Evans (1998) impose two assumptions: I β2 = 0: Sibling sex composition does not directly affect parents’ labor supply. I β3 = 0: The effect of family size on parents’ labor supply does not change with sibling sex composition. But as Angrist & Evans have noted, if parents change the family environment in response to child gender, then exogeneity of the same-sex instrument is violated. For example, if there is a son instead of a daughter, I the mother tends to work less (Rose 2000), the father more (Rose 2000; Lundberg & Rose 2003); I and the father are less likely to divorce or leave home (Dahl & Moretti 2008; Ananat & Michales 2008). Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Examples Examples I I I I So SameSex might affect parental labor supply via causal pathways other than M oreT han2. If so, Samesex would be an invalid instrument and the interaction between Samesex and M orethan2 should be taken into account. The pathways not through changing fertility choice are the “intensive margins.” We estimate the intensive margins using the 5% PUMS sample of all women from Angrist’s Data Archive, as a test for exogeneity of SameSex. N = 394, 840 We also test for significance of the interaction between Samesex and M orethan2. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Examples 1) Effects of Sex Composition on Fertility Choice Angrist-Evans (1998) 2SLS Model IV for Morethan2 Controls 1) First stage Fertility choice = Morethan2 Twin2nd Samesex Twin2nd × Samesex (1) (2) (3) (4) Samesex Twin2nd Twin2nd Twin2nd, Twin2nd×Samsex Boy1st, Boy2nd Boy1st, Boy2nd Samesex Samesex - 0.621 (0.008) - 0.622 (0.008) 0.059 (0.001) - 0.647 (0.011) 0.060 (0.001) -0.051 (0.015) 0.059 (0.001) - - Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Examples 2) Effects of Fertility Choice on Mothers’ Weeks Worked Angrist-Evans (1998) 2SLS Model IV for Morethan2 Controls 2) Second Stage Outcome= Weeks Worked Morethan2 Samesex Morethan2×Samesex (1) (2) (3) (4) Samesex Twin2nd Twin2nd Twin2nd, Twin2nd×Samsex Boy1st, Boy2nd Boy1st, Boy2nd Samesex Samesex -5.711 (1.155) - -3.663 (0.599) - - - -3.664 (0.598) -0.126 (0.077) - -4.138 (0.806) 2.15 (0.283) -5.213 (0.746) Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Examples Decompose the Sibling-Gender Effect on Mothers’ Weeks Worked Angrist-Evans (1998) (1) (2) (3) (4) Intensive margin, fixing n0 -0.338 (0.069) -0.338 (0.069) - -0.338 (0.069) -0.338 (0.069) - Intensive margin, fixing n1 - - Adjust for interaction - - -0.216 (0.036) -0.216 (0.036) -0.126 (0.077) -0.126 (0.077) - -0.338 (0.069) -0.338 (0.069) -0.342 (0.069) -0.245 (0.048) -0.553 (0.015) 0.211 (0.069) -0.100 (0.083) -0.308 (0.045) -0.346 (0.069) 3)Decompose the same-sex effect Extensive margin, given Samesex=0 Extensive margin, given Samesex=1 Total effect, given covariates Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Data I I Large and detailed data is a prerequisite for identifying the causal effects of sibling gender and family size. We use 2 national administrative data sets covering all of Taiwan: I I I I Birth Registry (1978-1999) University Entrance Test records (1996-2003) The Birth Registry was linked to University Entrance Test records by using children’s unique ID numbers. Education Outcomes: I I admitted to university at age 18, high school completion, which we use ”taking SAT tests” as a proxy for since most of graduating seniors take the tests. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Data I Analysis sample: I I First-borns with at least one sibling ⇒ We control for birth order Born prior to 1985, when access to ultrasound was limited. ⇒ We addressed the issue of endogenous child gender I I I I Sex ratio of first-borns = 1.039 - 1.042. Sex ratio of second-borns = 1.064 - 1.065. The F-statistic for the regression of the sex of second-born on family backgrounds was small and insignificant. Complete family size: We trace each mother who had the first baby between 1978 and 1984 for 15 to 22 years. No baby was born to the mothers in our sample in 1997-1998. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Characteristics of the Firstborn Sample Mean of firstborn singletons who have at least one sibling (part 1) Born in 1978-1984 Born in 1978-1979 Next sibling born by 1985 Sample size Sex-ratio of boys to girls Sex-ratio of the next sibling Birth years of the next siblings Complete family size Twins at 2nd birth Subject’s birth weight (kg) Urban (place of birth) Mother’s age at first birth Mother’s year of birth Father’s year of birth 833,371 1.045 1.070 1984 2.696 0.007 3.212 0.337 23.5 1958 1954 358,177 1.047 1.069 1981 2.784 0.006 3.223 0.340 23.2 1956 1952 336,828 1.046 1.068 1981 2.814 0.006 3.223 0.333 23.2 1956 1952 Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Characteristics of the Firstborn Sample Mean of firstborn singletons who have at least one sibling (part 2) Born in 1978-1984 5-yrs avg taxable income per capita in village Mother’s highest degree College/professional deg. or + HS diploma Vocational HS Junior HS Father’s highest degree College degree or above Professional degree HS diploma Vocational HS Junior HS Born in 1978-1979 All Next sibling born by 1985 728,791 726,922 723,397 0.069 0.061 0.187 0.261 0.064 0.053 0.161 0.214 0.061 0.052 0.158 0.214 0.063 0.073 0.092 0.181 0.237 0.061 0.067 0.088 0.160 0.187 0.058 0.066 0.086 0.160 0.187 Stacey H. Chen (Academia Sinica) Yen-Chien Chen for (Chinan with data support from Jin-Tan Liu (National Taiwan U.) Other covariates: dummies theU.)subject’s age, parental ages, and the Effects of Family Composition mother’s ageonatHuman firstCapital birth.Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Strong Demand for Sons Effect of Sibling Sex Composition on Sibsize All firstborns Two girls Mixed gender Born in urban Ln(taxable income per capital in village) Parental education Mother finished (1) (2) (3) junior HS or above 0.5376 (0.0024) 0.1007 (0.0018) -0.1782 (0.0017) 0.5375 (0.0024) 0.1008 (0.0018) -0.0757 (0.0021) -0.3717 (0.0044) 0.5363 (0.0024) 0.1 000 (0.0018) -0.0913 (0.0020) -0.2441 (0.0044) Yes 0.4098 (0.0035) 0.0613 (0.0025) -0.0714 (0.0029) -0.1574 (0.0059) Yes Sample size 833,371 833,371 833,371 264,105 R-squared adjusted 0.15 0.16 0.18 0.13 Sample size = 833,371. Other covariates: dummies for urban, parental ages and maternal age at first birth. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Instrument for Sibsize Robins and Greenland (1992), Cole and Hernan (2002) and VanderWeele (2010) have noted that with an endogenous mediator, identification of extensive and intensive margins cannot be achieved by using data on the triplet (Y, B, N ) alone. In this paper we introduce conventional Instrumental-Variable methods to the causal-inference literature by bring in a “fourth” variable, which is the instrument for the mediator. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Instrument for Sibsize Twins at the second birth (Rozensweig and Wolpin, 1980) I Potential issues: Compared to singletons, twins are lighter and live shorter on average. I After controlling for birth weight, college enrolment rates of twins and singletons are about the same (see figures 2a and 2b). I We include birthweight in X, as Rosenzweig and Zhang (2009) have suggested, to tackle the concern on the endowment deficit of twins. I Black, Devereux and Salvanes (2007) find similar results using data from Norway. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude 0 Probability of entering a college .05 .1 .15 .2 Figure 2a: Firstborn Girls’ College Enrolment Rate, Given Birth Weight 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 Birth weight Twin Singleton College outcome by birth weights(kg) for first−born girls Data: Taiwanese Birth Registry 1978-1984, firstborn girls Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude 0 Probability of entering a college .05 .1 .15 .2 Figure 2b: Firstborn Boys’ College Enrolment Rate, Given Birth Weight 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 Birth weight Twins Singleton College outcome by birth weights(kg) for first−born boys Data: Taiwanese Birth Registry 1978-1984, firstborn boys Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Sample Mean by Firstborn Gender (part 1) Firstborn girls Firstborn boys 407,467 425,904 Random Treatment Next sibling is male (Boy2nd=1) 0.52 0.52 Endogenous post-treatment More than 2 children (Morethan2=1) Sibsize 0.60 2.83 0.46 2.57 Instrument for fertility Twinning at 2nd birth 0.007 0.006 Outcome variables: Admitted to univ. High school completion 0.17 .24 0.15 .23 Sample size Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Sample Mean by Firstborn Gender (part 2) Control variables: Birth weight Born in urban Mother’s year of birth Mother’s age at first birth Father’s year of birth Taxable income per capita (1000) Mother’s highest degree College/professional degree or + HS diploma Vocational HS Junior HS Father’s highest degree College degree of above Professional degree HS diploma Vocational HS Junior HS Firstborn girls Firstborn boys 3.16 0.34 1957 23.49 1954 729.33 3.26 0.34 1958 23.46 1954 728.28 0.07 0.06 0.19 0.26 0.07 0.06 0.19 0.26 0.06 0.07 0.09 0.18 0.24 0.06 0.07 0.09 0.18 0.24 Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Outline of the main results Sample = Families with at least two children (1) First Stage (for endogenous fertility choice): I I I Dependent variable =M orethan2 or Sibsize Relevance of the instrument, Twinning at 2nd birth (T win2nd) Key covariate = having a 2nd-born brother (Boy2nd) (2) Second Stage (effects of fertility choice on child outcomes): I I I Education variables = Admitted to university, HS completion Effect of fertility choice on the fristborn’s education Effect of having Boy2nd on the firstborn’s education (3) Decompose the effect of sibling gender on education I I I Total effect Extensive and intensive margin Adjustment for interaction between potential sibsize and sibling gender. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude (1) First-Stage Estimates (Fertility Choice) More than two kids Firstborn girls Twinning at 2nd (Twin2nd) Next sibling is male (Boy2nd) Twin2nd× Boy2nd Birthweight Sibsize (1) (2) (3) (4) (5) 0.444 (0.008) -0.220 (0.001) 0.306 (0.013) -0.222 (0.001) 0.234 (0.017) 0.306 (0.013) -0.222 (0.001) 0.234 (0.017) -0.015 (0.002) 0.652 (0.015) -0.438 (0.002) 0.671 (0.023) -0.438 (0.002) -0.031 (0.030) -0.032 (0.003) N=407,467. Standard errors in (.). Other covariates include parental age, mother’s age at first birth, subject’s age, birthplace, urban dummy, and logarithm of taxable income per capita of birth village. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude (1) First-Stage Estimates (Fertility Choice) More than two kids Firstborn boys Twinning at 2nd (Twin2nd) Next sibling is male (Boy2nd) Twin2nd× Boy2nd Birthweight Sibsize (1) (2) (3) (4) (5) 0.566 (0.009) -0.064 (0.001) 0.530 (0.014) -0.064 (0.001) 0.065 (0.018) 0.530 (0.014) -0.064 (0.001) 0.065 (0.018) -0.023 (0.002) 0.723 (0.013) -0.102 (0.002) 0.721 (0.020) -0.102 (0.002) 0.005 (0.026) -0.039 (0.002) N=425,904. Standard errors in (.). Other covariates include parental age, mother’s age at first birth, subject’s age, birthplace, urban dummy, and logarithm of taxable income per capita of birth village. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude (2) Outcome equation (Admitted to University) OLS 2SLS Interact Firstborn girls More than 2 kids (Morethan2) Next sibling is male (Boy2nd) Interact Birthweight -0.0140 (0.0013) -0.0011 (0.0012) -0.0101 (0.0018) 0.0031 (0.0019) -0.0068 (0.0024) -0.0322 (0.0152) -0.0051 (0.0035) -0.0837 (0.0348) -0.0470 (0.0238) 0.0626 (0.0329) -0.0855 (0.0348) -0.0484 (0.0238) 0.0646 (0.0329) 0.0185 (0.0013) -0.0181 (0.0011) -0.0007 (0.0010) -0.0153 (0.0016) 0.0019 (0.0014) -0.0056 (0.0021) -0.0126 (0.0115) -0.0003 (0.0013) -0.0194 (0.0190) 0.0000 (0.0090) -0.0017 (0.0179) -0.0197 (0.0190) -0.0003 (0.0090) -0.0011 (0.0179) 0.0171 (0.0012) Morethan2×Boy2nd Birthweight Firstborn boys More than 2 kids (Morethan2) Next sibling is male (Boy2nd) Morethan2×Boy2nd Birthweight Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude (3) Decompose the effect on college education OLS Firstborn girls If next sibling is female: fertility channel AEM0 non-fertility channel AIM0 If next sibling is male: fertility channel AEM1 non-fertility channel AIM1 Interact Birthweight 0.0031 (0.0003) -0.0011 (0.0012) 0.0022 (0.0004) -0.0017 (0.0012) 0.0071 (0.0033) -0.0051 (0.0035) 0.0184 (0.0077) -0.0026 (0.0012) 0.0188 (0.0077) -0.0026 (0.0012) 0.0031 (0.0003) -0.0011 (0.0012) 0.0037 (0.0004) -0.0002 (0.0012) 0.0015 (0.0005) 0.0020 (0.0011) 0.0071 (0.0033) -0.0051 (0.0035) 0.0046 (0.0006) -0.0164 (0.0078) -0.0138 (0.0072) 0.0022 (0.0011) 0.0046 (0.0006) -0.0169 (0.0077) -0.0142 (0.0072) 0.0021 (0.0011) Adjust for interaction Total effect 2SLS Interact 0.0020 (0.0011) 0.0020 (0.0011) Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude (3) Decompose the effect on college education OLS Firstborn boys If next sibling is female: fertility channel AEM0 non-fertility channel AIM0 If next sibling is male: fertility channel AEM1 non-fertility channel AIM1 Interact. Birthweight 0.0012 (0.0001) -0.0007 (0.0010) 0.0010 (0.0001) -0.0009 (0.0011) 0.0008 (0.0007) -0.0003 (0.0013) 0.0012 (0.0012) -0.0009 (0.0011) 0.0013 (0.0012) -0.0009 (0.0011) 0.0012 (0.0001) -0.0007 (0.0010) 0.0013 (0.0001) -0.0005 (0.0011) 0.0004 (0.0001) 0.0005 (0.0010) 0.0008 (0.0007) -0.0003 (0.0013) 0.0013 (0.0001) -0.0008 (0.0016) 0.0001 (0.0011) 0.0005 (0.0010) 0.0013 (0.0001) -0.0008 (0.0016) 0.0001 (0.0011) 0.0005 (0.0010) Adjust for interaction Total effect 2SLS Interact. 0.0005 (0.0010) 0.0005 (0.0010) Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude 2SLS Estimates of the Effects on Education Mother has JHS diploma Firstborn girls Admitted to univ. HS completion Admitted to univ. HS completion Outcome mean 0.173 0.241 0.306 0.414 -0.0855 (0.0348) -0.0484 (0.0238) 0.0646 (0.0329) -0.1084 (0.0386) -0.0621 (0.0264) 0.0810 (0.0365) -0.1327 (0.0493) -0.0645 (0.0269) 0.1037 (0.0474) -0.1998 (0.0523) -0.0990 (0.0285) 0.1638 (0.0503) 407,467 407,467 129,287 129,287 More than 2 kids (Morethan2) Next sibling is male (Boy2nd) Morethan2×Boy2nd Sample size Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Decomposition Mother has JHS diploma Firstborn girls If next sibling is female: fertility channel AEM0 non-fertility channel AIM0 If next sibling is male: fertility channel AEM1 non-fertility channel AIM1 Adjust for interaction Total effect Admitted to univ. HS completion Admitted to univ. HS completion 0.0188 (0.0077) -0.0026 (0.0012) 0.0239 (0.0085) -0.0047 (0.0014) 0.0320 (0.0119) -0.0068 (0.0027) 0.0481 (0.0126) -0.0079 (0.0029) 0.0046 (0.0006) -0.0169 (0.0077) -0.0142 (0.0072) 0.0020 (0.0011) 0.0060 (0.0006) -0.0225 (0.0086) -0.0178 (0.0080) 0.0013 (0.0013) 0.0070 (0.0010) -0.0318 (0.0121) -0.0250 (0.0114) 0.0002 (0.0025) 0.0087 (0.0011) -0.0474 (0.0128) -0.0395 (0.0121) 0.0007 (0.0027) Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Summary of the Main Results (1/2) I I The total effects of having a brother on firstborn girls’/boys’ university entry are marginal and insignificant, because the extensive and intensive margins typically have opposite signs cancelling each other. Results are very sensitive to inclusion of the interaction between sibsize and sibling gender. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Summary of the Main Results (2/2) I I For firstborn boys, all margins are nearly zero. For firstborn girls: I I I I I Extensive margins (via reduced sibsize) are significantly positive, while intensive margins are negatively small through marginally significant. If the interaction is ignored, we understand the extensive margins. If the second born is female, it’s likely to be a larger family. The effect of having a brother via changing the sibsize is large. If the second born is male, it’s likely to be a smaller family. The effect of having a brother via changing the sibsize is small. But in smaller families, the direct effect of having a brother (i.e., intensive margin) is larger for girls. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Concluding remarks I I I I I We clarified the parameters of interest, pointing out the post-intervention bias in the previous estimators. We used instrumental-variable methods to address the endogenous post-treatment variable (family size). We applied the new method to examine the identification assumption of Angrist and Evans (1998). We constructed a unique administrative database minimally affected by sex-selective abortion. We introduced new outcomes variables – university attainment and high school completion – to measure the long-term impact of having a brother on human capital formation. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins Introduction The Problem Identification Data Results Conclude Concluding remarks I I We pointed to alternative pathways - intensive margin, less emphasized in the previous literature, for sibling gender to have an impact, namely, parental behavioural changes that alter family environments. We examined Goodkind’s (1996) hypothesis in a careful empirical design. I I I Goodkind hypothesized that pre-natal sex-selective abortion substitutes post-natal sex discrimination. Although pre-natal sex-selective abortion was not legally permitted during the sample period, parents in Taiwan have freely implemented their pro-male biased fertility-stopping rule because of no One-Child policy. The total effect of having a brother relative to a sister is close to zero because the positive effect due to the extensive margin cancel out a negative effect due to the intensive margin and a negative effect due to a decrease in the extensive margin. Stacey H. Chen (Academia Sinica) Yen-Chien Chen (Chinan U.) with data support from Jin-Tan Liu (National Taiwan U.) Effects of Family Composition on Human Capital Formation: Extensive and Intensive Margins