LATE and the Generalized Roy Model: Some Relationships James J. - - PowerPoint PPT Presentation

LATE and the Generalized Roy Model: Some Relationships

James J. Heckman University of Chicago Extract from: Building Bridges Between Structural and Program Evaluation Approaches to Evaluating Policy James J. Heckman (JEL 2010) Econ 312, Spring 2019

Heckman LATE and the Roy Model

Defining LATE

Heckman LATE and the Roy Model

• Question:* Derive the MTE from the sample selection model.

What parameters are identified by the selection model that are not identified by MTE? Explain the advantages and disadvantages of each approach. *Answer after reading these slides

Heckman LATE and the Roy Model

LATE

• LATE is defined by the variation of an instrument.
• The instrument in LATE plays the role of a randomized

assignment.

• Randomized assignment is an instrument.
• Y0 and Y1 are potential ex-post outcomes.
• Instrument Z assumes values in Z, z ∈ Z.

Heckman LATE and the Roy Model

• D(z): indicator of hypothetical choice representing what choice

the individual would have made had the individual’s Z been exogenously set to z.

• D(z) = 1 if the person chooses (is assigned to) 1.
• D(z) = 0, otherwise.
• One can think of the values of z as fixed by an experiment or

by some other mechanism independent of (Y0, Y1).

• All policies are assumed to operate through their effects on Z.
• It is assumed that Z can be varied conditional on X.

Heckman LATE and the Roy Model

• Three assumptions define LATE.

(IA-1) (IA-1)

(Y0, Y1, {D(z)}z∈Z) ⊥ ⊥ Z | X

(IA-2) (IA-2)

Pr(D = 1 | Z = z) is a nontrivial function of z conditional on X.

Heckman LATE and the Roy Model

(IA-3) (IA-3)

For any two values of Z, say Z = z1 and Z = z2, either D(z1) ≥ D(z2) for all persons, or D(z1) ≤ D(z2) for all persons.

• This condition is a statement across people.
• This condition does not require that for any other two values of

Z, say z3 and z4, the direction of the inequalities on D(z3) and D(z4) have to be ordered in the same direction as they are for D(z1) and D(z2).

• It only requires that the direction of the inequalities are the

same across people.

• Thus for any person, D(z) need not be monotonic in z.

Heckman LATE and the Roy Model

• Under LATE conditions, for two distinct values of Z, z1 and z2,

IV applied to LATE(z2, z1) = E(Y1 − Y0 | D(z2) = 1, D(z1) = 0), if the change from z1 to z2 induces people into the program (D(z2) ≥ D(z1)).

• This is the mean return to participation in the program for

people induced to switch treatment status by the change from z1 to z2.

Heckman LATE and the Roy Model

• LATE does not identify which people are induced to change

their treatment status by the change in the instrument.

• It leaves unanswered many policy questions.
• For example, if a proposed program changes the same

components of vector Z as used to identify LATE but at different values of Z (say z4, z3), LATE(z2,z1) does not identify LATE(z4, z3).

Heckman LATE and the Roy Model

• If the policy operates on different components of Z than are

used to identify LATE, one cannot safely use LATE to identify marginal returns to the policy.

• It does not, in general, identify treatment on the treated, ATE
• r a variety of criteria.
• But using the implicit economics of the problem one can do

better as I show below.

Heckman LATE and the Roy Model

Identifying Policy Parameters Y1 = µ1(X)+U1, Y0 = µ0(X)+U0, C = µC(Z)+UC, (1)

• (X, Z) are observed by the analyst.
• U0, U1, UC are unobserved.

Heckman LATE and the Roy Model

• Define Z to include all of X.
• Variables in Z not in X are instruments.
• ID = E(Y1 − Y0 − C | I) = µD(Z) − V

µD(Z) = E(µ1(X) − µ0(X) − µC(Z) | I) V = −E(U1 − U0 − UC | I).

• Choice equation:

D = 1(µD(Z) > V ). (2)

• In the early literature that implemented this approach µ0(X),

µ1(X), and µC(Z) were assumed to be linear in the parameters, and the unobservables were assumed to be normal and distributed independently of X and Z.

Heckman LATE and the Roy Model

• The essential aspect of the structural approach is joint

modeling of outcome and choice equations.

• Structural econometricians have developed nonparametric

identification analyses for the Roy and generalized Roy models.

Heckman LATE and the Roy Model

A useful fact: Assume Z ⊥ ⊥ V Then Choice Probability : P(z) = Pr(D = 1 | Z = z) = Pr(µD(z) ≥ V ) = Pr µD(z) σV ≥ V σV

• P(z) = F

V σV

• µD(z)

σV

• UD = F

V σV

• V

σV

• ;

Uniform(0, 1)

Heckman LATE and the Roy Model

P(z) = Pr

• F V

σV

µD(z) σV

• ≥ F

V σV

• V

σV

• = Pr (P(z) ≥ UD)

P(z) is the p(z)th quantile of UD.

Heckman LATE and the Roy Model

Recall Y = DY1 + (1 − D)Y0 = Y0 + D(Y1 − Y0) Keep X implicit (condition on X = x) E(Y | Z = z) = E(Y0) + E(Y1 − Y0 | D = 1, Z = z)P(z)

• from law of iterated expectations

= E(Y0) + E(Y1 − Y0 | P(z) ≥ UD)P(z) ∴ It depends on Z only through P(Z). E(Y | Z = z′) = E(Y0) + E(Y1 − Y0 | P(z′) ≥ UD)P(z′)

Heckman LATE and the Roy Model

• What is E(Y1 − Y0 | P(z) ≥ UD)? (Treatment on the treated)
• Assume (Y1, Y0, UD) (absolutely) continuous.
• The joint density of (Y1 − Y0, UD): fY1−Y0,UD(y1 − y0, uD).
• Does not depend on Z.
• It may, in general, depend on X.
• E(Y1 − Y0 | P(z) ≥ UD)

=

∞

• −∞

P(z)

• (y1 − y0)fy1−y0,uD(y1 − y0, uD) duDd(y1 − y0)

Pr(P(z) ≥ UD)

Heckman LATE and the Roy Model

• Recall that

UD = F

V σV

• V

σV

• .
• UD is a quantile of the V /σV distribution.

Heckman LATE and the Roy Model

• By construction, UD is Uniform(0, 1) (this is the definition of a

quantile).

• ∴ fUD(uD) = 1.
• Also, Pr(P(z) ≥ UD) = P(z).
• By law of conditional probability,

fY1−Y0,UD(y1 − y0, uD) = fY1−Y0,UD(y1 − y0 | UD = uD) fUD(uD)

=1

.

Heckman LATE and the Roy Model

E(Y1 − Y0 | P(z) ≥ UD) =

P(z)

• ∞
• −∞

(y1 − y0)fY1−Y0,UD(y1 − y0, uD) d(y1 − y0) duD P(z) E(Y1 − Y0 | P(z) ≥ UD) =

P(z)

• ∞
• −∞

(y1 − y0)fY1−Y0,UD(y1 − y0 | UD = uD) d(y1 − y0) duD P(z) =

P(z)

• E(Y1 − Y0 | UD = uD) duD

P(z)

Heckman LATE and the Roy Model

∴ E(Y | Z = z) = E(Y0) +

P(z)

• E(Y1 − Y0 | UD = uD)duD

∂E(Y | Z = z) ∂P(z) = E(Y1 − Y0 | UD = P(z))

• marginal gains for

people with UD=P(z)

= MTE(UD) for UD = P(Z) E(Y | Z = z′) = E(Y0) +

P(z′)

• E(Y1 − Y0 | UD = uD)duD

Heckman LATE and the Roy Model

• Suppose P(z) > P(z′)

∴ E(Y | Z = z) − E(Y | Z = z′) = =

P(z)

• P(z′)

E(Y1 − Y0 | UD = uD)duD = E(Y1 − Y0 | P(z) ≥ UD ≥ P(z′)) Pr(P(z) ≥ UD ≥ P(z′))

Heckman LATE and the Roy Model

Notice Pr(P(z) ≥ UD ≥ P(z′)) =

P(z)

• P(z′)

duD = P(z) − P(z′) E(Y | Z = z) − E(Y | Z = z′) = E(Y1 − Y0 | P(z) ≥ UD ≥ P(z′))

• LATE

(P(z) − P(z′))

Heckman LATE and the Roy Model

E(Y | Z = z) − E(Y | Z = z′) P(z) − P(z′) = LATE(z, z′) =

P(z)

• P(z′)

MTE(uD)duD P(z) − P(z′)

Heckman LATE and the Roy Model

• Question: In what sense is E(Y1 − Y0 | P(z) ≥ UD) a

measure of surplus of agents for whom P(z) ≥ UD?

Heckman LATE and the Roy Model

Appendix: The Generalized Roy Model for the Normal Case

Heckman LATE and the Roy Model

Y1 = µ1(X) + U1 Y0 = µ0(X) + U0 C = µC(Z) + UC Net Benefit: I = Y1 − Y0 − C I = µ1(X) − µ0(X) − µC(Z)

• µD(Z)

+ U1 − U0 − UC

• −V

(U0, U1, UC) ⊥ ⊥ (X, Z) E(U0, U1, UC) = (0, 0, 0) V ⊥ ⊥ (X, Z)

Heckman LATE and the Roy Model

• Assume normally distributed errors.
• Assume Z contains X but may contain other variables

(exclusions) Y = DY1 + (1 − D)Y0

• bserved Y

D = 1(I ≥ 0) = 1(µD(Z) ≥ V )

• Assume V ∼ N(0, σ2

V )

Heckman LATE and the Roy Model

• Propensity Score:

Pr(D = 1 | Z = z) = Φ µD(z) σV

• E(Y | D = 1, X = x, Z = z) = µ1(X) + E(U1 | µD(z) ≥ V )
• K1(P(z))

because (X, Z) ⊥ ⊥ (U1, V ).

• Under normality we obtain

E

• U1
• µD(z)

σV ≥ V σV

• =

Cov(U1, V

σV )

Var( V

σV )

˜ λ µD(z) σV

• Heckman

LATE and the Roy Model

• Why?

U1 = Cov

• U1, V

σV V σV + ε1 ε1 ⊥ ⊥ V E V σV | µD(z) σV ≥ V σV

• =

µD (z) σV

• −∞

t

1 √ 2πe

−t2 2 dt µD (z) σV

• −∞

1 √ 2πe

−t2 2 dt

= ˜ λ µD(z) σV

• =

−1 √ 2πe(− 1

2)

µD (z)

σV

2

Φ

• µD(z)

σV

• = ˜

λ µD(z) σV

• =

−φ

• µD(z)

σV

• Φ
• µD(z)

σV

• Heckman

LATE and the Roy Model

• Notice

lim

µD(z)→∞

˜ λ µD(z) σV

• =0

lim

µD(z)→−∞

˜ λ µD(z) σV

• = − ∞
• Propensity score:

P(z) = Pr(D = 1 | Z = z) = Φ µD(z) σV

• ∴

µD(z) σV

• = Φ−1 (Pr(D = 1 | Z = z))

Heckman LATE and the Roy Model

• Thus we can replace µD(z)

σV

with a known function of P(z)

Heckman LATE and the Roy Model

• Notice that because (X, Z) ⊥

⊥ (U, V ), Z enters the model (conditional on X) only through P(Z).

• This is called index sufficiency.

Heckman LATE and the Roy Model

• Put all of these results together to obtain

E (Y | D = 1, X = x, Z = z) = µ1(x) +

• Cov(U1, V

σV )

Var( V

σV )

• ˜

λ µD(z) σV

• = E (Y1 | D = 1, X = x, Z = z) = µ1(x) +
• Cov(U1, V

σV )

Var( V

σV )

• ˜

λ µD(z) σV

• ˜

λ(z) = E V σV | V σV < µD(z) σV

• < 0

λ(z) = E V σV | V σV ≥ µD(z) σV

• > 0

E (Y | D = 0, X = x, Z = z) = µ0(x) +

• Cov(U0, V

σV )

Var( V

σV )

• λ

µD(z) σV

• Var

V σV

• = 1

Heckman LATE and the Roy Model

V σV = −(U1 − U0 − UC) σV Cov

• U1, V

σV

• = − Cov
• U1, V

σV

• + Cov
• U0, V

σV

• + Cov
• UC, V

σV

• In Roy model case (UC = 0),

Cov

• U1, V

σV

• = − Cov
• U1, U1 − U0

σV

• = −Cov (U1 − U0, U1)
• Var(U1 − U0)

Heckman LATE and the Roy Model

• We can identify µ1(x), µ0(x)
• From Discrete Choice model we can identify

µD(z) σV = µ1(x) − µ0(x) − µC(z) σV

• If we have a regressor in X that does not affect µC(z) (say

regressor xj, so ∂µC (z)

∂xj

= 0), we can identify σV and µC(z).

• ∴ We can identify the net benefit function and the cost

function up to scale.

• ∴ We can compute ex-ante subjective net gains.

Heckman LATE and the Roy Model

• Method generalizes: Don’t need normality
• Need “Large Support” assumption to identify ATE and TT
• E (Y | D = 1, X = x, Z = z) = µ1(x) +

control function

• K1(P(z))

E (Y | D = 0, X = x, Z = z) = µ0(x) + K0(P(z))

• control function

lim

P(z)→1 E (Y | D = 1, X = x, Z = z) = µ1(x)

lim

P(z)→0 E (Y | D = 0, X = x, Z = z) = µ0(x)

Heckman LATE and the Roy Model

• If we have this condition satisfied, we can identify ATE

E(Y1 − Y0 | X = x) = µ1(x) − µ0(x)

• ATE is defined in a limit set. This is true for any model with

selection on unobservables (IV; selection models)

Heckman LATE and the Roy Model

• What about treatment on the treated?

E(Y1 − Y0 | D = 1, X = x, Z = z)

Heckman LATE and the Roy Model

(a) From the data, we observe

E(Y1 | D = 1, X = x, Z = z)

(b) Can also create it from the model (c) E(Y0 | D = 1, X = x, Z = z) is a counterfactual

We know E(Y0 | D = 0, X = x, Z = z) = µ0(x) + Cov

• U0, V

σV

• λ

µD(Z) σV

• (this is data)

Heckman LATE and the Roy Model

(d) We seek

E(Y0 | D = 1, X = x, Z = z) = µ0(x) + Cov

• U0, V

σV

• ˜

λ µD(z) σV

• But under normality, we know Cov
• U0, V

σV

• We know µD(Z)

σV

• ˜

λ(·) is a known function.

• Can form ˜

λ

• µD(z)

σV

• and can construct counterfactual.

Heckman LATE and the Roy Model

• More generally, without normality but with (X, Z) ⊥

⊥ (U, V ),

E(Y1 | D = 1, X, Z) = E(Y | D = 1, X = x, Z = z) = µ1(x) + K1(P(z)) E(Y0 | D = 0, X, Z) = E(Y | D = 0, X = x, Z = z) = µ0(x) + ˜ K0(P(z)) where K1(P(z)) = E(U1 | D = 1, X = x, Z = z) = E

• U1 | µD(z)

σV > V σV

• ˜

K1(P(z)) = E

• U1 | µD(z)

σV > V σV

• ˜

K0(P(z)) = E

• U0 | µD(z)

σV > V σV

• Heckman

LATE and the Roy Model

• Use the transformation

FV σV µD(z) σV

• = P(z)

FV σV V σV

• = UD

(a uniform random variable) D = 1 µD(z) σV ≥ V σV

• = 1 (P(z) ≥ UD)

K1(P(z)) = E(U1 | P(z) > UD) K1(P(z))P(z) + ˜ K1(P(z))(1 − P(z)) = 0 ∴ we can construct ˜ K1(P(z))

Heckman LATE and the Roy Model

• Symmetrically

˜ K0(P(z)) = E(U0 | P(z) ≤ UD) K0(P(z)) = E(U0 | P(z) > UD) (1 − P(z)) ˜ K0(P(z)) + P(z)K0(P(z)) = 0

Heckman LATE and the Roy Model

• ∴ If we have “identification at infinity,” we can construct

E(Y1 − Y0 | X = x) = µ1(x) − µ0(x)

• We can construct TT

E(Y1 − Y0 | D = 1, X = x, Z = z) = = [µ1(x) + K1(P(z))]

• factual

− [µ0(x) + K0(P(z))]

• counterfactual
• But we can form µ1(x) + K1(P(z)) from data
• We get µ0(x) from limit set P(z) → 0 identifies µ0(x)
• We can form K0(P(z)) = − ˜

K0(P(z))

P(z) 1−P(z)

• ∴ Can construct the desired counterfactual mean.

Heckman LATE and the Roy Model

• Notice how we can get Effect of Treatment for People at the Margin:

E(Y1 − Y0 | I = 0, X = x, Z = z)

• Under normality we have (as a result of independence and normality)

E(Y1 − Y0 | I = 0, X = x, Z = z) = µ1(x) − µ0(x) + E

• U1 − U0 | µD(z)

σV = V σV , X = x, Z = z

• = µ1(x) − µ0(x) + Cov
• U1 − U0, V

σV µD(z) σV In the Roy model case where UC = 0 but µC(z) = 0 = µ1(x) − µ0(x) − σV µD(z) σV

• = µ1(x) − µ0(x) − µD(z)

= µC(z) (marginal gain = marginal cost)

Heckman LATE and the Roy Model

• MTE is

E(Y1 − Y0 | V = v, X = x, Z = z) = = µ1(x) − µ0(x) + Cov

• U1 − U0, V

σV

• v
• Effect of Treatment for People at the Margin picks v = µD(z)

σV

• Notice we can use the result that

µD(z) σV = F −1

• V

σV

(P(z))

V = F −1

• V

σV

(UD)

Heckman LATE and the Roy Model

• Effect of Treatment for People at Margin of Indifference

Between Taking Treatment and Not: E(Y1 − Y0 | I = 0, X = x, Z = z) = = µ1(x) − µ0(x) + Cov

• U1 − U0, V

σV

• F −1
• V

σV

(P(z))

• MTE:

E(Y1 − Y0 | UD = uD, X = x, Z = z) = = µ1(x) − µ0(x) + Cov

• U1 − U0, V

σV

• F −1
• V

σV

(UD)

Heckman LATE and the Roy Model

• Recent (1987 and Later!) Advances in Econometrics:

(a) Relax normality (b) Do not assume linearity of µ1(X) and µ0(X) in terms of X (c) Do not require identification at infinity but only because they

abandon pursuit of ATE, TT, TUT or else assume that (Y1, Y0) ⊥ ⊥ D | X (matching assumption)

(d) Identification at infinity in some version or the other is required

for ATE, TT, TUT as long as there is selection on unobservables (i.e., (Y1, Y0) ⊥

• ⊥ D | X)

Heckman LATE and the Roy Model

End of Example of Normal Model

Heckman LATE and the Roy Model

Appendix: Nonparametric Identification of the Roy Model

• (Y0, Y1) potential outcomes
• I ∗ = Y1 − Y0 choice index
• Observe Y1 if Y1 ≥ Y0.
• Observe Y0 if Y1 < Y0.
• Cannot simultaneously observe Y0 and Y1.
• Generalized Roy model: I = Y1 − Y0 − C.
• C depends on Z.

Heckman LATE and the Roy Model

• Heuristically, we can conduct an identification analysis

assuming we know I = I ∗ σY1−Y0 = Y1 − Y0 σY1−Y0 for each person where D = 1(I > 0).

• See Cosslett (1983), Manski (1988), Matzkin (1992).
• Assumes there is an an instrument Z that shifts C.
• Even though we do not ever observe I, we observe (Y0, D) and

(Y1, D). We never observe the full triple (Y0, Y1, D) for anyone.

• We only observe some components of C.

Heckman LATE and the Roy Model

• Under conditions specified in the literature, F(Y0, I|X, Z) and

F(Y1, I|X, Z) are identified where Y0 = µ0(X) + U0 E(Y0 | X) = µ0(X) Y1 = µ1(X) + U1 E(Y1 | X) = µ1(X) I ∗ = µI(X, Z) + UI I = µI (X,Z)

σUI

+ UI

σUI

• Source: Heckman (1990), Heckman and Honor´

e (1990).

• The key idea in these papers is “sufficient” variation in Z

holding X fixed.

Heckman LATE and the Roy Model

Sketch of the Proof

• From the left-hand side of

Pr(D = 1|X, Z) = Pr(µI(X, Z) + UI ≥ 0|X, Z), we can identify the distribution of

UI σUI , as well as µI (X,Z) σUI

.

• This is true under normality or any assumed form for the

distribution of

UI σUI .

• It is also true more generally.
• One does not have to assume the distribution of UI is known or

that the functional form of µI(X, Z) is linear, e.g., µI(X, Z) = XβI + Zγ.

• See the conditions in the Matzkin (1992) paper and the survey

in Matzkin, 2007, Handbook of Econometrics.

Heckman LATE and the Roy Model

• This more general claim requires full support of Z and

restrictions on µI(X, Z). See the “Matzkin conditions” in Cunha, Heckman, and Navarro (2007, IER).

• A key condition is

Support µI(X, Z) σUI

• ⊇ Support

UI σUI

• and other regularity conditions.
• Commonly it is assumed that for a fixed X

Support µI(X, Z) σUI

• = (−∞, ∞).
• This is called “identification at infinity.” When we vary Z over

its conditional support (for each X) we trace out the full support of

UI σUI .

Heckman LATE and the Roy Model

Identifying the Joint Distribution of (Y0, I)

• From data, we know the conditional distribution of Y0:

F(Y0 | D = 0, X, Z) = Pr(Y0 ≤ y0 | µI(X, Z) + UI ≤ 0, X, Z)

• Multiply this by Pr(D = 0 | X, Z):

F(Y0 | D = 0, X, Z) Pr(D = 0 | X, Z) = Pr(Y0 ≤ y0, I ∗ ≤ 0 | X, Z) (*)

• Follow the analysis of Heckman (1990), Heckman and Smith

(1998), and Carneiro, Hansen, and Heckman (2003).

Heckman LATE and the Roy Model

• Left hand side of (*) is known from the data.
• Right hand side:

Pr

• Y0 ≤ y0, UI

σUI < −µI(X, Z) σUI | X, Z

• Since we know µI(X, Z)

σUI from the previous analysis, we can vary it for each fixed X.

Heckman LATE and the Roy Model

• If µI(X, Z) gets small (µI(X, Z) → −∞), recover the marginal

distribution Y and in this limit set we can identify the marginal distribution of Y0 = µ0(X) + U0 ∴ can identify µ0(X) in limit.

• (See Heckman, 1990, and Heckman and Vytlacil, 2007.)
• More generally, we can form:

Pr

• U0 ≤ y0 − µ0(X), UI

σUI ≤ −µI(X, Z) σUI | X, Z

• X and Z can be varied and y0 is a number. We can trace out

joint distribution of

• U0, UI

σUI

• by varying (Y0, Z) for each fixed

X.

Heckman LATE and the Roy Model

• ∴ Recover joint distribution of

(Y0, I) =

• µ0(X) + U0, µI(X, Z) + UI

σUI

• .
• Three key ingredients:

1 The independence of (U0, UI) and (X, Z). 2 The assumption that we can set µI(X, Z)

σUI to be very small (so we get the marginal distribution of Y0 and hence µ0(X)).

3 The assumption that µI(X, Z)

σUI can be varied independently of µ0(X).

• Trace out the joint distribution of
• U0, UI

σUI

• . Result generalizes

easily to the vector case. (Carneiro, Hansen, and Heckman, 2003, IER; Heckman and Vytlacil, Part I).

Heckman LATE and the Roy Model

• Another way to see this is to write:

F(Y0 | D = 0, X, Z) Pr(D = 0 | X, Z)

• This is a function of µ0(X) and µI(X, Z)

σUI (Index sufficiency)

Heckman LATE and the Roy Model

• Varying the µ0(X) and µI(X, Z)

σUI traces out the distribution of

• U0, UI

σUI

• .
• Effectively we observe the pairs
• I

σUI , Y1

• and
• I

σUI , Y0

• .
• We never observe the triple
• I

σUI , Y0, Y1

• .

Heckman LATE and the Roy Model

• Use the intuition that we “know” I.
• Actually we observe

F(Y0 | I < 0, X, Z) and F(Y1 | I ≥ 0, X, Z) and Pr(I ≥ 0 | X, Z)

• Can construct the joint distributions F(Y0, I | X, Z) and

F(Y1, I | X, Z).

Heckman LATE and the Roy Model

Roy Case

• Armed with normality (or the nonparametric assumptions in

Heckman and Honor´ e, 1990), we can estimate Cov(I, Y1) = Var(Y1) − Cov(Y0, Y1) σ2

Y1 + σ2 Y0 − 2σY1,Y0

Cov(I, Y0) = −Var(Y0) − Cov(Y0, Y1) σ2

Y1 + σ2 Y0 − 2σY1,Y0

.

• We know Var Y1, Var Y0 (e.g. normal selection model or use

limit sets).

• ∴ Cov(Y0, Y1) is identified (actually over-identified).
• This line of argument does not generalize if we add a cost

component (C) that is unobserved (or partly so).

• It carries through exactly if C(Z) is solely a function of Z.

Heckman LATE and the Roy Model

Intuition

• In the Roy model the decision rule is generated solely by

(Y1, Y0).

• Knowing agent choices we observe the relative order (and

magnitude) of Y1 and Y0.

• Thus we get a second valuable piece of information from agent
• choices. This information is ignored in statistical approaches to

program evaluation.

• But does this analysis generalize?

Heckman LATE and the Roy Model

Generalized Roy Model

• Add cost

I = Y1 − Y0 − C

• Assume that we do not directly observe C.

Observe Y1 | I > 0, Observe Y0 | I < 0, I = Y1 − Y0 − C

• Var(Y1 − Y0 − C)

.

Heckman LATE and the Roy Model

• We can identify Var Y1 and can identify Var Y0.
• But we cannot directly identify Cov(Y0, Y1) which measures

comparative advantage in Willis-Rosen model.

• Notice, however, we can determine if

E(Y1 | I > 0) > E(Y1) E(Y0 | I < 0) > E(Y0)

• (Are people who work in a sector above average for the sector?)

Heckman LATE and the Roy Model