Introduction

Peiyu He

Peking university

June 2, 2021

Peiyu He (Peking university) Short title June 2, 2021 1 / 11

Setting and notation

Population UN consisting of N units. From UN , n units are randomly

sampled into the experimental sample. Of the n sampled units, m
units are randomly assigned to the treatment condition.
The indicator variable XiT be one if unit i is assigned to the treatment
condition, and let the indicator XiC be one if unit i is assigned to the
control condition. If XiT = XiC = 0, then the unit is unsampled.
Potential outcome for each units: y0i and y1i . The analyst then
observes y0i when XiC = 1 and y1i when XiT = 1.
Assume an index ordering i = 1, . . . , N such thatX1T , . . . , XmT = 1,
C , . . . , X C = 1, and the remaining N − n unsampled units comes
Xm+1 n
last.

Peiyu He (Peking university) Short title June 2, 2021 2 / 11

Setting and notation

Given elements vi , wi for i = 1, . . . , N, define

PN
finite population mean µN (v ) = N1 i=1 vi
PN 2
finite population variance σN2 (v ) = N1 i=1 {vi − µN (v )}
finite population
PNcovariance
σN (v , w ) = N1 i=1 {vi − µN (v )} {wi − µN (w )}
Then the average treatment effect for the population UN is

τN = µN (y1 ) − µN (y0 )

The difference-in-means estimator of τN is

m n
1 X 1 X
τ̂N = µ̂N (y1 ) − µ̂N (y0 ) = y1i − y0i
m n−m
i=1 i=m+1

Peiyu He (Peking university) Short title June 2, 2021 3 / 11

Problems
Since EX (τ̂N ) = τN , our inferential target is the variance of τˆN :
 
1 N −m 2 N − (n − m) 2
VN = σN (y1 ) + σN (y0 ) + 2σN (y1 , y0 )
N −1 m n−m
.
We have unbiased estimators of σN 2 (y ) and σ 2 (y ):
1 N 0

m
N −1 X
2
σ̂N (y1 ) = {y1i − µ̂N (y1 )}2
N(m − 1)
i=1
,
n
N −1 X
2
σ̂N (y0 ) = {y0i − µ̂N (y0 )}2
N(n − m − 1)
i=m+1

But both potential outcomes y0i and y1i for the same unit can never
be observed simultaneously, σN (y1 , y0 ) is unidentifiable when N is
finite. We need turn to construct the bounds of the variance.
Peiyu He (Peking university) Short title June 2, 2021 4 / 11
Neyman approximations

When n = N, the sampling variance of the difference-in-means

estimator reduces to
 
1 n−m 2 m 2
Vn = σn (y1 ) + σ (y0 ) + 2σn (y1 , y0 )
n−1 m n−m n

Neyman proposed estimators of Vn that uses the inequality

1/2
2σN (y1 , y0 ) ≤ 2 σn2 (y1 ) σn2 (y0 ) ≤ σn2 (y1 ) + σn2 (y0 )


An upper bound estimate for Vn is obtained by setting

2σN (y1 , y0 ) = σn2 (y1 ) + σn2 (y0 ):
 2
σ̂n (y1 ) σ̂n2 (y0 )

a n
V̂n = +
n−1 m n−m

V̂na is conservative and is known to be unbiased for Vn when effects

are constant.

Peiyu He (Peking university) Short title June 2, 2021 5 / 11

Neyman approximations

Neyman also proposed a method for computing sharpest bounds on

Vn given only knowledge of the second moments σn2 (y1 ) , σn2 (y0 ):
 
b± 1 n−m 2 m 2
 2 2
1/2
V̂n = σ̂n (y1 ) + σ̂ (y0 ) ± 2 σ̂n (y1 ) σ̂n (y0 )
n−1 m n−m n

Neyman recommended choosing V̂nb+ as a conservative approximation

to the true variance.
However, under the setting considered, we have stronger information
than the second moment, that is, the marginal distribution of y1 , y0 ,
they can be used to obtain asymptotically sharp bounds on VN .

Peiyu He (Peking university) Short title June 2, 2021 6 / 11

Sharp bounds on σN (y1 , y0 ) given marginal distributions of
outcomes
Let GN (y ) = N −1 N I (y1i ≤ y ), FN (y ) = N −1 N
P P
Z 1 i=1 i=1 I (y0i ≤ y )
H
σN (y1 , y0 ) = GN−1 (u)FN−1 (u)du − µN (y1 ) µN (y0 )
0
Z 1
L
σN (y1 , y0 ) = GN−1 (u)FN−1 (1 − u)du − µN (y1 ) µN (y0 )
0

Theorem (1)
Given only GN and FN and no other information on the joint distribution
of (y1 , y0 ), the bound
L H
σN (y1 , y0 ) ≤ σN (y1 , y0 ) ≤ σN (y1 , y0 )

is sharp. The upper bound is attained if (y1 , y0 ) ∼ GN−1 (U), FN−1 (U) for


a uniform random
 −1 variable U on [0, 1], The lower bound is attained if
−1
(y1 , y0 ) ∼ GN (U), FN (1 − U) .
Peiyu He (Peking university) Short title June 2, 2021 7 / 11
Proof
Let HN (y1 , y0 ) be the joint distribution function of (y1 , y0 ) and
consider two other distributions HNH (y1 , y0 ) = min {GN (y1 ) , FN (y0 )},
HNL (y1 , y0 ) = max {0, GN (y1 ) + FN (y0 ) − 1}
All three distributions have the same marginals GN and FN , according
to a result by Hoeffding, recounted in Tchen, we have:

EH L (y1 y0 ) ≤ EHN (y1 y0 ) ≤ EH H (y1 y0 )

N N

Since GN−1 (U), FN−1 (U) ∼ HNH and GN−1 (U), FN−1 (1 − U) ∼ HNL .
 

Z 1
EH H (y1 y0 ) = GN−1 (u)FN−1 (u)du
N
0
Z 1
EH L (y1 y0 ) = GN−1 (u)FN−1 (1 − u)du
N
0
complete the proof.
Peiyu He (Peking university) Short title June 2, 2021 8 / 11
Sharp bounds on VN given marginal distributions of
outcomes
 L H
Theorem 1 implies
 VN , VN is the sharpest interval bound for VN : 
1 N − m 2 N − (n − m) 2
VNH = σN (y1 ) + H
σN (y0 ) + 2σN (y1 , y0 )
N −1 m n−m
 
L 1 N −m 2 N − (n − m) 2 L
VN = σN (y1 ) + σN (y0 ) + 2σN (y1 , y0 )
N −1 m n−m
In practice,
PN we have estimates for GN and FN : ĜN P(yN) = C
−1 T −1
m i=1 Xi I (y1i ≤ y ) , F̂N (y ) = (n − m) i=1 Xi I (y0i ≤ y )
h i
And then we obtain the interval estimator V̂NL , V̂NH for VN :
 
H 1 N −m 2 N − (n − m) 2 H
V̂N = σ̂N (y1 ) + σ̂N (y0 ) + 2σ̂N (y1 , y0 )
N −1 m n−m
 
L 1 N −m 2 N − (n − m) 2 L
V̂N = σ̂N (y1 ) + σ̂N (y0 ) + 2σ̂N (y1 , y0 )
N −1 m n−m

Peiyu He (Peking university) Short title June 2, 2021 9 / 11

Asymptotic sharpness of interval estimator

Theorem (2)
Suppose the following conditions hold as N → ∞:
1. (m/N, n/N) → (θρ, θ) for θ ∈ (0, 1] and ρ ∈ (0, 1);
2. HN converges weakly to a limit distribution H with marginals G (y ) =
H(y , ∞) and F (y ) = H(∞, y ) ;
3. GN (y ) → G (y ) at any discontinuity point of G , and FN (y ) → F (y ) at
any discontinuity point of F;
4. The sequences of distributions represented by {GN }N and {FN }N are
uniformly square-integrable. That is, as β → ∞,
   
1 X N  1 X N 
sup y1i2 , sup y0i2 → 0
N N 2
 N N 2

i:y1i ≥β i:y0i ≥β

Peiyu He (Peking university) Short title June 2, 2021 10 / 11

Asymptotic sharpness of interval estimator

Theorem (2)
Then for the collection H of all bivariate distributions with marginals G
and F , the moments of each h ∈ H exist up to second order and

1 − θρ 1 − θ(1 − ρ)
NVNH → VarH (y1 ) + VarH (y0 ) + 2 sup Covh (y1 , y0 ) ,
θρ θ(1 − ρ) h∈H
1 − θρ 1 − θ(1 − ρ)
NVNL → VarH (y1 ) + VarH (y0 ) + 2 inf Covh (y1 , y0 ) .
θρ θ(1 − ρ) h∈H
 
Moreover V̂NH − VNH , V̂NL − VNL = oP (1/N).

Peiyu He (Peking university) Short title June 2, 2021 11 / 11