Professional Documents
Culture Documents
Kosuke Imai
Harvard University
Fall 2019
Fig. 1 Model sensitivity of ATE estimates for imbalanced raw and balanced matched data. Thi
figure presents an artificial data set of treated units represented by ‘‘T’’ and control units represented
by ‘‘C.’’Kosuke
The Imai
vertical axis plots Yi and the horizontal
(Harvard) axis plots Xi. The
Matching Methods panels depictFall
Stat186/Gov2002 estimates
2019 3of
/ 18the
Selection Bias
Assumptions
1 Overlap: 0 < Pr(Ti = 1 | Xi = x) < 1 for any x
2 Ignorability: {Yi (1), Yi (0)}⊥
⊥Ti | Xi = x for any x
Bias decomposition (Heckman et al. 1998. Econometrica ):
E(Yi (0) | Ti = 1) − E(Yi | Ti = 0)
Z
= E(Yi (0) | Ti = 1, Xi = x)dFXi |Ti =1 (x)
S1 \S
Z
− E(Yi (0) | Ti = 0, Xi = x)dFXi |Ti =0 (x)
S0 \S
| {z }
bias due to lack of common support
Z
+ E(Yi (0) | Ti = 0, Xi = x)d{FXi |Ti =1 (x) − FXi |Ti =0 (x)}
|S {z }
bias due to imbalance of observables
Z
+ {E(Yi (0) | Ti = 1, Xi = x) − E(Yi (0) | Ti = 0, Xi = x)}dFXi |Ti =1 (x)
S
| {z }
bias due to unobservables
F
e (Xi | Ti = 1) = F
e (Xi | Ti = 0)
No model dependence
But, infeasible when
covariate is continuous
there are many covariates
D(Xi , Xj ) = |π(X
[i ) − π(X i = 1 | Xi ) − Pr(Tj = 1 | Xj )|
[j )| = |Pr(T\ \
D(Xi , Xj ) = |logit(π(X
[i )) − logit(π(X
[j ))|
Common matching methods (Rubin. 2006. Matched Sampling for Causal
Effects. Cambridge University Press; Stuart. 2010. Stat. Sci.):
one-to-one, one-to-many
caliper
with and without replacement
optimal matching (Rosenbaum 1991. J. Am. Stat. Assoc)
full matching (Rosenbaum. 1989. J. Royal Stat. Soc. B; Hansen. 2004. J. Am.
Stat. Assoc)
Kosuke Imai (Harvard) Matching Methods Stat186/Gov2002 Fall 2019 6 / 18
Propensity Score as a Balancing Score (Rosenbaum and Rubin.
1983. Biometrika)
π(Xi ) = Pr(Ti = 1 | Xi )
Balancing property:
Ti ⊥
⊥ Xi | π(Xi )
1
Pn
1
T X X
P
n1 i=1 i ij − |Mi | i ∈Mi
0 i 0j
Conservative Candidates
● Matched
Aristocrat ● ● ● Unmatched
University: Degree ● ●
University: Oxbridge ● ●
Schooling: Regular ● ●
Schooling: Public ● ●
Schooling: Eton ● ●
Miner ●
Journalist ● ●
Union Official ●
White Collar ● ●
Business ● ●
Local Politician ● ●
Civil Servant ● ●
Doctor ● ●
Solicitor ● ●
Barrister ● ●
Teacher ● ●
Female ● ●
Year of Death ● ●
Year of Birth ● ●
Standardized Bias
Reject H0
Reject H0 Accept H0 Reject H0 Accept H0 Accept H0
of a
of no difference of no difference of no difference of a difference difference of a difference
α 2 α 2
t−stat t−stat
Figure 1: Tests of equivalence versus tests of difference. The left panel depicts the logic of tests of
Inverting the test the largest equivalence region which is
difference under the null hypothesis of no difference. The right panel depicts the logic of tests of
consistent with the data at the (1 − α) × 100% confidence level
equivalence under the null hypothesis of difference.
There are three factors that can result in the t-statistic lying in either the tails or the center of the
t Kosuke
distribution. If the mean difference betweenMatching
Imai (Harvard) the two populations
Methods is small, then the t-statistic will
Stat186/Gov2002 Fall 2019 11 / 18
FIGURE 2 Results of Equivalence Tests
Effect of Ethnic Quota on Redistribution
Observed Equivalence Tests Equivalenc
Mean Sci. Rev.)
(Dunning and Nilekani. 2013. Am. Political Confidence
Difference Equivalence Range: +/- 0.36 σ Interval (+/-
Variable (Scale of Var) (Scale of V
# Illiterates −257.6 # Illiterates 6 33.9
# Marginal Workers −12.5 # Marginal Workers 83.4
# Households −99 # Households 209.2
Ordered list determining
Agricultural Laborers −14.1 Agricultural Laborers 93.8
female nonworkers
Population (0−6 ) have −107 Population (0−6 ) 231.8
Percentage SC 0 Percentage SC 0
Percentage ST 0 Percentage ST 0
Note: The
Kosuke Imai (Harvard) observed mean difference is the mean of the treated
Matching Methods group minus
Stat186/Gov2002 Fallthe mean 12
2019 of / 18
the c
Bias of Matching
Bias of matching arises because of imbalance:
1 X
B(Xi , XMi ) = E(Yi (0) | Ti = 1, Xi ) − E Yi 0 XMi
|Mi | 0
i ∈Mi
V(τ̂match | X, T)
n n
1 X 1 X
≈ V(Yi (1) | X, T) + W 2 · V(Yi (0) | X, T)
n12 i:T =1 n02 i:T =0 i
i i
n 2
X Ti Wi
= + (1 − Ti ) V(Yi | X, T)
n1 n0
i=1
Unobserved confounders:
an error has been made, of an old kind, in arguing from cor-
relation to causation, ... the possibility should be explored that
the different smoking classes, non-smokers, cigarette smok-
ers, cigar smokers, pipe smokers, etc., have adopted their
habits partly by reason of their personal temperaments and
dispositions, and are not lightly to be assumed to be equiva-
lent in their genotypic composition. (Fisher. 1958. Nature)
36,975 heavy smokers paired with nonsomokers based on age,
race, education, marital status, various health history measures,
etc. (Hammond. 1964. J. Natl. Cancer Inst.)
Of these pairs, 122 pairs had exactly one person died of lung
cancer – 110 heavy smokers
Sensitivity analysis based on McNemar’s test (maximum p-value):
< 0.0001 (Γ = 3), 0.004(Γ = 4), 0.03(Γ = 5), 0.1(Γ = 6)
Recommended readings:
Ho et al. 2007. “Matching as Nonparametric Preprocessing for
Reducing Model Dependence in Parametric Causal Inference.”
Political Analysis
Stuart. 2010. “Matching methods for causal inference: A review
and a look forward.” Statistical Science
Imbens and Rubin. Chapters 12–15, 17–19, and 22.