Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
How to Make Causal Inferences with Time-Series Cross-Sectional Data
Search
Matthew Blackwell
April 13, 2013
Science
2
180
How to Make Causal Inferences with Time-Series Cross-Sectional Data
Matthew Blackwell
April 13, 2013
Tweet
Share
Other Decks in Science
See All in Science
ABEMAの効果検証事例〜効果の異質性を考える〜
s1ok69oo
3
1.5k
Non-Gaussian methods for causal discovery
sshimizu2006
0
180
論文輪読会 第16回 "NeRF:Representing Scenes as Neural"
academix
0
460
Microbiology Labs.
maleehafatima
0
130
Introduction to Graph Neural Networks
joisino
4
1.4k
Презентация программы бакалавриата СПбГУ "Искусственный интеллект и наука о данных"
dscs
0
100
BMI 研究はなぜ同じ失敗を繰り返すのか(日本BMI研究会, 2021.11.5)
ykamit
1
1.9k
「国と音楽」 ~spotifyrを用いて~ #muana
bob3bob3
2
330
Ph.D. defense "Convex Manifold Approximation for Tensors"
gkazunii
0
180
Machine Learning for Materials (Lecture 3)
aronwalsh
0
830
Unlocking Healthcare data: the power of Open Formats in Python Data Science
whitone
0
150
Endocannabinology 101
drbonci
PRO
0
250
Featured
See All Featured
A designer walks into a library…
pauljervisheath
199
23k
Large-scale JavaScript Application Architecture
addyosmani
504
110k
VelocityConf: Rendering Performance Case Studies
addyosmani
320
23k
Scaling GitHub
holman
457
140k
Building Flexible Design Systems
yeseniaperezcruz
318
37k
Side Projects
sachag
451
41k
Product Roadmaps are Hard
iamctodd
43
9.7k
Designing Experiences People Love
moore
136
23k
The Success of Rails: Ensuring Growth for the Next 100 Years
eileencodes
30
6k
Testing 201, or: Great Expectations
jmmastey
27
6.3k
Keith and Marios Guide to Fast Websites
keithpitt
408
22k
How To Stay Up To Date on Web Technology
chriscoyier
782
250k
Transcript
How to Make Causal Inferences with Time-Series Cross-Sectional Data Matthew
Blackwell University of Rochester Adam Glynn Harvard University
How to Make Causal Inferences with Time-Series Cross-Sectional Data
How to Make Causal Inferences with Time-Series Cross-Sectional Data Very
Carefully.
How to Make Causal Inferences with Time-Series Cross-Sectional Data Using
weights.
ۺ˞ ۢ˞ ۹˞ ۢ ۹ ۺ
What is the effect of A on Y? ۺ˞ ۢ˞
۹˞ ۢ ۹ ۺ
What is the effect of A on Y? contemporaneous ۺ˞
ۢ˞ ۹˞ ۢ ۹ ۺ
What is the effect of A on Y? treatment history
ۺ˞ ۢ˞ ۹˞ ۢ ۹ ۺ
Shouldn't we have more notation?
ۢ ۢ Ɛ ۢ Treatment history
Shouldn't we have more notation?
ۢ ۢ Ɛ ۢ Treatment history
Shouldn't we have more notation? Specific instance of a treatment history ۼ ۼ Ɛ ۼ
ۢ ۢ Ɛ ۢ Treatment history
ۺ ۼ Potential outcomes Shouldn't we have more notation? Specific instance of a treatment history ۼ ۼ Ɛ ۼ
The effect of history
The effect of history ণ ۼ ۼƓ
ۦ=ۺ ۼ ˞ ۺ ۼƓ ? Average Treatment History Effect
The effect of history ণ ۼ ۼƓ
ۦ=ۺ ۼ ˞ ۺ ۼƓ ? Average Treatment History Effect ATHE
The effect of history ণ ۼ ۼƓ
ۦ=ۺ ۼ ˞ ۺ ۼƓ ? Average Treatment History Effect 1 1 1 1 1 1 1 ATHE
The effect of history ণ ۼ ۼƓ
ۦ=ۺ ۼ ˞ ۺ ۼƓ ? Average Treatment History Effect 1 1 1 1 1 1 1 0 0 0 0 0 0 0 vs ATHE
The effect of history
The effect of history Blip Effect ণ۽ ۼ ˞
ۦ=ۺ ۼ ˞ ˞ ۺ ۼ ˞ ?
The effect of history 1 0 0 0 0 0
0 0 vs 0 0 0 0 0 0 Blip Effect ণ۽ ۼ ˞ ۦ=ۺ ۼ ˞ ˞ ۺ ۼ ˞ ?
The effect of history 1 0 0 0 0 0
0 0 vs 0 0 0 0 0 0 Blip Effect ণ۽ ۼ ˞ ۦ=ۺ ۼ ˞ ˞ ۺ ۼ ˞ ?
The effect of history 1 0 0 0 0 vs
0 0 0 1 1 1 Blip Effect ণ۽ ۼ ˞ ۦ=ۺ ۼ ˞ ˞ ۺ ۼ ˞ ? 1 1 1
The effect of history 1 0 vs 1 1 1
Blip Effect ণ۽ ۼ ˞ ۦ=ۺ ۼ ˞ ˞ ۺ ۼ ˞ ? 1 1 1 1 1 1 1 1 1
The effect of history
The effect of history Contemporaneous Effect of Treatment ণ
ۦ=ণ۽ ۼ ˞ ?
The effect of history Contemporaneous Effect of Treatment ণ
ۦ=ণ۽ ۼ ˞ ? CET
The effect of history 1 0 vs Contemporaneous Effect of
Treatment ণ ۦ=ণ۽ ۼ ˞ ? CET
The effect of history 1 0 vs Contemporaneous Effect of
Treatment ণ ۦ=ণ۽ ۼ ˞ ? CET Marginalize over the past
TSCS data under sequential ignorability Treatment is unrelated to the
potential outcomes ...conditional on the covariate history. ۺ ۼ е е ۢ ^۹ ۺ ˞ ۢ ܃˞ ۼ ˞
How conditioning leads you astray
How conditioning leads you astray ...for some questions.
How conditioning leads you astray ...for some questions. ۺ
૿ ۢ ଁ ۹ ଂ ۺ˞ ଃ ۢ˞
ۺ˞ ۢ˞ ۢ ۹ ۺ How conditioning leads you astray
...for some questions. ۺ ૿ ۢ ଁ ۹ ଂ ۺ˞ ଃ ۢ˞
ۺ˞ ۢ˞ ۢ ۹ ۺ How conditioning leads you astray
...for some questions. We “fix” these ۺ ૿ ۢ ଁ ۹ ଂ ۺ˞ ଃ ۢ˞
ۺ˞ ۢ˞ ۢ ۹ ۺ How conditioning leads you astray
...for some questions. We “fix” these ۺ ૿ ۢ ଁ ۹ ଂ ۺ˞ ଃ ۢ˞
ۺ˞ ۢ˞ ۢ ۹ ۺ How conditioning leads you astray
We “fix” these ...for some questions. ۺ ૿ ۢ ଁ ۹ ଂ ۺ˞ ଃ ۢ˞
ۺ˞ ۢ˞ ۢ ۹ ۺ How conditioning leads you astray
...for some questions. ۺ ૿ ۢ ଁ ۹ ଂ ۺ˞ ଃ ۢ˞
ۺ˞ ۢ˞ ۢ ۹ ۺ How conditioning leads you astray
এ૾ ...for some questions. ۺ ૿ ۢ ଁ ۹ ଂ ۺ˞ ଃ ۢ˞
ۺ˞ ۢ˞ ۢ ۹ ۺ How conditioning leads you astray
এ૾ ? ? ? ? ? ? ...for some questions. ۺ ૿ ۢ ଁ ۹ ଂ ۺ˞ ଃ ۢ˞
ۺ˞ ۢ˞ ۢ ۹ ۺ How conditioning leads you astray
এ૾ ? ? ? ? ? ? ...for some questions. CET: (1,0) vs (0,0) ATHE: (0,1) vs (0,0) ATHE: (1,1) vs (0,0) এ૾ ̪ এଁ ̪ এ૾ এଁ ۺ ૿ ۢ ଁ ۹ ଂ ۺ˞ ଃ ۢ˞
How weighting can help
ۺ˞ ۢ˞ ۢ ۹ ۺ How weighting can help
ۺ˞ ۢ˞ ۢ ۹ ۺ How weighting can help ۸܃
ಿ ܍ 2T=ۢ܃܍ ^ۢ˞ ۹ ۺ˞ ?
ۺ˞ ۢ˞ ۢ ۹ ۺ How weighting can help ۸܃
ಿ ܍ 2T=ۢ܃܍ ^ۢ˞ ۹ ۺ˞ ? We weight to create balance
ۺ˞ ۢ˞ ۢ ۹ ۺ How weighting can help We
weight to create balance ۸܃ ಿ ܍ 2T=ۢ܃܍ ^ۢ˞ ۹ ۺ˞ ?
ۺ˞ ۢ˞ ۢ ۹ ۺ How weighting can help ۸܃
ಿ ܍ 2T=ۢ܃܍ ^ۢ˞ ۹ ۺ˞ ? Unconfounded No posttreatment bias
How weighting can help
How weighting can help ۦ=ۺ ۼ ۼ˞ ?
ۦ۸ =ۺ ^ۢ ۼ ۢ˞ ۼ˞ ? ૿ ۼ ଁ ۼ˞
How weighting can help ۦ=ۺ ۼ ۼ˞ ?
ۦ۸ =ۺ ^ۢ ۼ ۢ˞ ۼ˞ ? ૿ ۼ ଁ ۼ˞ WLS
How weighting can help ۦ=ۺ ۼ ۼ˞ ?
ۦ۸ =ۺ ^ۢ ۼ ۢ˞ ۼ˞ ? ૿ ۼ ଁ ۼ˞ WLS CET: (1,0) vs (0,0) ATHE: (0,1) vs (0,0) ATHE: (1,1) vs (0,0) এ૾ ଁ ଁ
The Long Arm of the Democratic Peace?
The Long Arm of the Democratic Peace? Democracy in year
t War in year t
The Long Arm of the Democratic Peace? Democracy in year
t War in year t Democratic Peace Literature
The Long Arm of the Democratic Peace? Democracy in year
t War in year t Democratic Peace Literature History of Democracy
The Long Arm of the Democratic Peace? Democracy in year
t War in year t Democratic Peace Literature History of Democracy Can we estimate this?
%FQFOEFOU WBSJBCMF %JTQVUF #,5 .JTTQFDJĕFE *158 .PEFM $VNVMBUJWF .PEFM .4.
%FNPDSBDZ #MJQ ˞૿ଅˣˣˣ ૿ଅ૿ $VNVMBUJWF %FNPDSBDZ ˞૿૿૿ ˞૿૿ଃˣˣˣ ૿૿ଁ ૿૿ଂ (SPXUI ˞ଂଇଂଆˣˣˣ ˞ଃଂଅ૿ˣˣˣ 0CTFSWBUJPOT ଁ૿ ଃଃଇ ଁ૿ ଃଃଇ ଁ૿ ଃଃଇ /PUF ˣQ ˣˣQ ˣˣˣQ Revisiting Beck, Katz, and Tucker (1998)
%FQFOEFOU WBSJBCMF %JTQVUF #,5 .JTTQFDJĕFE *158 .PEFM $VNVMBUJWF .PEFM .4.
%FNPDSBDZ #MJQ ˞૿ଅˣˣˣ ૿ଅ૿ $VNVMBUJWF %FNPDSBDZ ˞૿૿૿ ˞૿૿ଃˣˣˣ ૿૿ଁ ૿૿ଂ (SPXUI ˞ଂଇଂଆˣˣˣ ˞ଃଂଅ૿ˣˣˣ 0CTFSWBUJPOT ଁ૿ ଃଃଇ ଁ૿ ଃଃଇ ଁ૿ ଃଃଇ /PUF ˣQ ˣˣQ ˣˣˣQ Revisiting Beck, Katz, and Tucker (1998)
Democracy in year t War in year t Economic Growth
in year t History of Democracy Misspecification of an ATHE Time-Varying Confounder
%FQFOEFOU WBSJBCMF %JTQVUF #,5 .JTTQFDJĕFE *158 .PEFM $VNVMBUJWF .PEFM .4.
%FNPDSBDZ #MJQ ˞૿ଅˣˣˣ ૿ଅ૿ $VNVMBUJWF %FNPDSBDZ ˞૿૿૿ ˞૿૿ଃˣˣˣ ૿૿ଁ ૿૿ଂ (SPXUI ˞ଂଇଂଆˣˣˣ ˞ଃଂଅ૿ˣˣˣ 0CTFSWBUJPOT ଁ૿ ଃଃଇ ଁ૿ ଃଃଇ ଁ૿ ଃଃଇ /PUF ˣQ ˣˣQ ˣˣˣQ Revisiting Beck, Katz, and Tucker (1998)
TSCS data under unmeasured confounding
TSCS data under unmeasured confounding ۺ܃ ۼ е е
ۢ܃ ^۹ ܃ ۢ ܃˞ ۼ ˞ ۶
TSCS data under unmeasured confounding Treatment is unrelated to the
potential outcomes ۺ܃ ۼ е е ۢ܃ ^۹ ܃ ۢ ܃˞ ۼ ˞ ۶
TSCS data under unmeasured confounding Treatment is unrelated to the
potential outcomes ...conditional on the covariate history ۺ܃ ۼ е е ۢ܃ ^۹ ܃ ۢ ܃˞ ۼ ˞ ۶
TSCS data under unmeasured confounding Treatment is unrelated to the
potential outcomes ...conditional on the covariate history ۺ܃ ۼ е е ۢ܃ ^۹ ܃ ۢ ܃˞ ۼ ˞ ۶ ...and a time-fixed unmeasured confounder.
How unit-specific weighting can help
How unit-specific weighting can help ۺ˞ ۢ˞ ۢ ۹ ۺ
۶
How unit-specific weighting can help ۺ˞ ۢ˞ ۢ ۹ ۺ
۶ ۸܃ ಿ ܍ 2T=ۢ܃܍ ^ۢ˞ ۹ ۺ˞ ۶?
How unit-specific weighting can help ۺ˞ ۢ˞ ۢ ۹ ۺ
۶ ۸܃ ಿ ܍ 2T=ۢ܃܍ ^ۢ˞ ۹ ۺ˞ ۶? Weighting balances the treatment groups.
ۺ˞ ۢ˞ ۢ ۹ ۺ How unit-specific weighting can help
۶ ۸܃ ಿ ܍ 2T=ۢ܃܍ ^ۢ˞ ۹ ۺ˞ ۶?
A weighting approach to fixed effects
A weighting approach to fixed effects 1 Estimate unit-specific probability
of treatment over time and construct weights.
A weighting approach to fixed effects 1 Estimate unit-specific probability
of treatment over time and construct weights. 2 Estimate a pooled outcome model with unit-specific weights
k-order sequential ignorability
k-order sequential ignorability ۺ܃ ۼ е е ۢ܃ ^۹
܃̂˞܅ ۢ ܃˞̂˞܅ ۼ ˞̂˞܅ ۶
k-order sequential ignorability Only the last k periods matter. ۺ܃
ۼ е е ۢ܃ ^۹ ܃̂˞܅ ۢ ܃˞̂˞܅ ۼ ˞̂˞܅ ۶
Blip effect: (1,0) vs (0,0) Time periods Blip effect 10
25 50 75 100 125 0.2 0.3 0.4 0.5 0.6 0.7
Blip effect: (1,0) vs (0,0) Time periods Blip effect 10
25 50 75 100 125 0.2 0.3 0.4 0.5 0.6 0.7
Pooled Blip effect: (1,0) vs (0,0) Time periods Blip effect
10 25 50 75 100 125 0.2 0.3 0.4 0.5 0.6 0.7 • • • • • •
Pooled Outcome fixed effects Blip effect: (1,0) vs (0,0) Time
periods Blip effect 10 25 50 75 100 125 0.2 0.3 0.4 0.5 0.6 0.7 • • • • • •
Pooled Outcome fixed effects Blip effect: (1,0) vs (0,0) IPTW
true weights Time periods Blip effect 10 25 50 75 100 125 0.2 0.3 0.4 0.5 0.6 0.7 • • • • • •
Pooled IPTW fixed effects Outcome fixed effects IPTW true weights
Blip effect: (1,0) vs (0,0) Time periods Blip effect 10 25 50 75 100 125 0.2 0.3 0.4 0.5 0.6 0.7 • • • • • •
Treatment History Effect: (1,1) vs (0,0) Time periods ATHE 10
25 50 75 100 125 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
Treatment History Effect: (1,1) vs (0,0) Time periods ATHE 10
25 50 75 100 125 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 Time periods ATHE 10 25 50 75 100 125 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
Pooled Treatment History Effect: (1,1) vs (0,0) Time periods ATHE
10 25 50 75 100 125 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 • • • • • •
Pooled Outcome fixed effects Treatment History Effect: (1,1) vs (0,0)
Time periods ATHE 10 25 50 75 100 125 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 • • • • • •
Pooled Outcome fixed effects IPTW true weights Treatment History Effect:
(1,1) vs (0,0) Time periods ATHE 10 25 50 75 100 125 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 • • • • • •
Pooled IPTW fixed effects Outcome fixed effects IPTW true weights
Treatment History Effect: (1,1) vs (0,0) Time periods ATHE 10 25 50 75 100 125 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 • • • • • •
How to make causal inferences with TSCS data
How to make causal inferences with TSCS data Very carefully
How to make causal inferences with TSCS data Very carefully
Even under strong assumptions, conditional estimators cannot recover ATHEs.
How to make causal inferences with TSCS data Very carefully
Using weights Even under strong assumptions, conditional estimators cannot recover ATHEs.
How to make causal inferences with TSCS data Very carefully
Using weights Even under strong assumptions, conditional estimators cannot recover ATHEs. A fixed effects weighting approach can recover ATHEs and CETs even with unmeasured confounding.