Professional Documents
Culture Documents
Chapter on Matching
This is the process that describes selection; there are two components:
Note these are all potential outcomes and treatments, so the missing observation problem means
that from each of these sets, only one is observed.
There are in principle four types of units
These can by represented by the following typology characterizing units by treatment and as-
signment types:
1
L7 EC 402, 2022
Z=0
Di0
=0 Di0 = 1
Di1 = 0 Never-taker Defier
Z=1
Di1 =1 Complier Always-taker
We rule out defiers or deniers by assumption; this is termed the monotonicity assumption.
They do the opposite of the assignment rule.
This leaves only the compliers, for whom we are able to observe outcomes with and without
participation.
1.1 Assumptions
1. SUTVA: Potential outcomes and treatments of unit i are independent of the potential assign-
ments, treatment and outcomes of unit j ̸= i. I.e.
Di (Z) = Di (Zi )
Yi (D, Z) = Yi (Di , Zi ).
2. Ignorability or random assignment, or “as if” random assignment. All units have the same
probability of assignment to treatment. I.e. P (Zi = 1) = P (Zj = 1)
In other words, the probability of treatment is different across the two assignment groups.
4. Exclusion restriction: Assignment (second argument) affects outcome only through treatment:
Using lower case z to denote variable (and upper case Z for its realization), the observed
outcome can be written as a function of potential outcomes:
2
L7 EC 402, 2022
5. Monotonicity: Di1 − Di0 ≥ 0 ∀ i, or vice-versa That is, no one does the opposite of their
assignment.
The last three assumptions together are known as strong monotonicity. In a purely mechanical
sense what is getting ruled out can be seen by the following tabulation where the assumption
is Di1 − Di0 ≥ 0
Di1 Di0 Di1 − Di0 Assump met? Type
1 1 0 Yes Always-taker
1 0 1 Yes Complier
0 1 -1 No Defier
0 0 0 Yes Never-taker
In the context of impact assessment, the RHS (X) variable is the binary participation variable D,
which can be endogneous in the presence of partial compliance. A natural choice for the IV Z
is the assignment rule, which is distinct from the participation decision D unless there is perfect
compliance.
Under the assumptions in section 1.1, what treatment effects can be identified?
Causal ITT effect of Z on D (in other words, on participation): E(Di1 − Di0 )
Causal effect of D on Y : E(Yi1 − Yi0 ) = E[Yi (1, z) − Yi (0, z)] But this a LATE, and not an
ATT, as seen below.
Cov(Di Zi )
E(Di |Zi = 1) − E(Di |Zi = 0) =
V ar(Zi )
= E(Di |Zi = 1) − E(Di0 |Zi = 0)
1
This means that the “first stage” regression of D on Z yields the causal effect of assignment
on treatment. This correlation should be non-zero by above assumptions. In other words the
assignment either improves or worsens the probability of participation.
3
L7 EC 402, 2022
Cov(Yi Zi )
⇒ E(Yi |Zi = 1) − E(Yi |Zi = 0) =
V ar(Zi )
= E[Yi (Di , 1)|Zi = 1] − E[Yi (Di0 , 0)|Zi = 0]
1
To examine the role of assumptions in greater detail, consider in tabular form the causal effects of
Z on Y at the unit level:
Z=0
Di0
=0 Di0 = 1
Di1 = 0 Never-taker Defier
Yi (0, 1) - Yi (0, 0) Yi (0, 1) - Yi (1, 0)
Z=1
Di1 = 1 Complier Always-taker
Yi (1, 1) − Yi (0, 0) Yi (1, 1) − Yi (1, 0)
Z=0
Di0
=0 Di0 = 1
Di1 = 0 Never-taker Defier
Yi (0, 1) - Yi (0, 0) Yi (0, 1) - Yi (1, 0)
= Yi0 − Yi0 = 0 = −(Yi1 − Yi0 )
Z=1 1
Di = 1 Complier Always-taker
Yi (1, 1) − Yi (0, 0) Yi (1, 1) − Yi (1, 0)
= Yi1 − Yi0 = Yi1 − Yi1 = 0
By SUTVA and random assignment, the causal effect can be written separately for each i and
for each cell
Montonicity implies that there are no defiers and there are at least some compliers
4
L7 EC 402, 2022
LATE, for reasons outlined above. A formal statement of Local Average Treatment Effect (LATE)
theorem is
E[Yi |Zi = 1] − E[Yi |Zi = 0]
β̂IV = W ald = = E[Yi1 − Yi0 |Di1 > Di0 ]
E[Di |Zi = 1] − E[Di |Zi = 0]
(The LHS is a Wald Estimator). Proof of LATE theorem: The numerator:
E[Yi |Zi = 1] = E[Yi0 + (Yi1 − Yi0 )Di |Zi = 1] = E[Yi0 + (Yi1 − Yi0 )Di1 ]
This is a consequence of SUTVA, as seen above. Similarly, the second term in the numerator
E[Yi |Zi = 0] = E[Yi0 + (Yi1 − Yi0 )Di |Zi = 0] = E[Yi0 + (Yi1 − Yi0 )Di0 ]
E[(Yi1 − Yi0 )(Di1 − Di0 )] = E[(Yi1 − Yi0 ).1|(Di1 > Di0 )]P (Di1 > Di0 )
The denominator, as seen earlier is E(Di1 − Di0 ). By monotonicity and similar argument again,
E(Di1 − Di0 ) = E[(Di1 − Di0 )|Di1 > Di0 ]P (Di1 > Di0 )
Thus treatment effects for compliers (the first term) may be cancelled out by treatment effects for
the defiers (second term) even though both sets of effects are positive.
Therefore the causal effect can only be estimated for the group of compliers. Therefore the IV
estimator identifies the LATE.
3 Encouragement designs
Recall the types of RCT designs. With the characterisation of compliance and noncompliance, and
the IV toolkit in hand, we can examine encouragement designs. By way of a hypothetical example:
say
5
L7 EC 402, 2022
the eligibility criterion is women aged 18 to 19 years. Absent any other interventions, say the
takeup rate is 40 percent.
the encouragement design would consist of the researchers offering an additional 30 percent
discount on the bicycle price, but this additional discount would be offered only to a randomly
drawn subset of the eligible women. Say the take up rate in this treated subgroup is 60 percent.
this means that the intervention led to a 20 percentage point increase in the take up of the
subsidy.
importantly, note that the “control” group also had access to the subsidy, but at the policy
rate of 50 percent.
The ITT estimates are based on the random allocation, but they are effectively the treatment
effects of the encouragement (the additional 30 percent discount), not of the programme
intervention (50 percent discount) per se.
But the LATE theorem above tells us we can recover the effect of the programme anyway.
The only price we pay is that it is defined only for the set of compliers, and with all the
assumptions necessary to identify it
Recall that the Wald estimator is the ITT divided by the proportion of compliers. How is the
percentage of compliers receiving the encouragement to be computed? Under exclusion and
monotonicity, the proportion of compliers is defined by the difference in takeup rate among
those (randomly selected into) receiving the encouragment and the takeup rate among the
(randomly selected) non-encouraged. Proof not required.
Newer work addresses sample size calculations for LATE (see e.g. Bansak, 2018)
The Fishman et al. (2021) paper uses an encouragement design. Their ITT estimates have the
additional wrinkle of including covariates, but that does not change the intuitions (or results)
derived above.