CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 1 /21
Estimating Causal Effects of Air QualityRegulations Using Principal Stratification
for Spatially-Correlated MultivariateIntermediate Outcomes
Corwin M. Zigler
Department of Biostatistics, Harvard School of Public Health
May 24, 2012
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 2 /21
Air Quality, Health, andRegulation
Accountability Assessment
• Long term exposure to air pollution is bad for health.• EPA estimates ≈ $25 billion per year on air quality
management.• 1970 Clean Air Act.
• For a specific regulatory action:• What were the causal effects on air quality?• What were the causal effects on health?
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 3 /21
1990 Clean Air ActAmendments (CAAA)
EPA designates counties as:
1 Attainment of air quality standards for PM10 .2 Nonattainment of air quality standards for PM10 :
• Required states to implement plans to achievestandards.
What were the causal effects of the 1990nonattainment designations for PM10 on:
• Pollution• Ambient concentrations of PM10 and O3 in 1999-2001.
• Health• All-cause Medicare mortality in 2001.
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 4 /21
Causal Inference for Air QualityRegulations
1 Potential outcomes in the EPA regulatory environment.
2 Air quality is a posttreatment concomitant variable.• ⇒ Principal stratification.
3 Regulations affect multiple pollutants.• ⇒ Multivariate continuous intermediate variable.
4 Pollution is spatially correlated.• Hierarchical spatial model.
5 Interference between observations (no SUTVA).
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 4 /21
Causal Inference for Air QualityRegulations
1 Potential outcomes in the EPA regulatory environment.2 Air quality is a posttreatment concomitant variable.
• ⇒ Principal stratification.
3 Regulations affect multiple pollutants.• ⇒ Multivariate continuous intermediate variable.
4 Pollution is spatially correlated.• Hierarchical spatial model.
5 Interference between observations (no SUTVA).
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 4 /21
Causal Inference for Air QualityRegulations
1 Potential outcomes in the EPA regulatory environment.2 Air quality is a posttreatment concomitant variable.
• ⇒ Principal stratification.3 Regulations affect multiple pollutants.
• ⇒ Multivariate continuous intermediate variable.
4 Pollution is spatially correlated.• Hierarchical spatial model.
5 Interference between observations (no SUTVA).
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 4 /21
Causal Inference for Air QualityRegulations
1 Potential outcomes in the EPA regulatory environment.2 Air quality is a posttreatment concomitant variable.
• ⇒ Principal stratification.3 Regulations affect multiple pollutants.
• ⇒ Multivariate continuous intermediate variable.4 Pollution is spatially correlated.
• Hierarchical spatial model.
5 Interference between observations (no SUTVA).
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 5 /21
Observed Regulation Programand Overall Causal Effect
Observed Regulation Program
●●
●
●
●
●●●●
●●●●●●●●●●●●●●
●●●●
●●
●●●●●●●●●●●●●
●
●
●
●●●●●● ●
●
●●●●
●
●●
●●
●●
●●
●
●●●
●
●
●●●●●●●●●●●●
●●●
●
●●
●
●
●
●
● ●
●●●●
●
● ●
●● ●●●
●●
●●
●●●●●●
●
●●
●● ●
●
●●●●
●●●
●●
●
●
●●●
●
●●●
●
●
●
●●
●●
●●●●●●●●●
●●
●
●●●●
●
●
●
●
●● ●●●●
●
●●●●
●●●
●
●●●●●●
●
●●●●●
●● ●●
●
●●
●● ●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●●
●
●●●
●
●●
●●
●●
●
●●
●
●●●
●
●●●●●●●●
●●●●●
●●●●●●●
●
●
●●
●● ●
● ●●●
●●
●●
●
●
●
●●
●
●●●
●●●●●
●●●●●
●●
●
● ●●●
●●●
●●●●●●
●
●
●
●●●
●●
●●
●●
●●
●●
●●●●●●
●●●●
●●●●●●●●●●●
●
●
●●
●●
●
●
●
●
●
●
AttainmentNonattainment
Observational unit:pollution monitor(point-referenced data)
Overall Causal Effect
●●
●
●
●
●●●●
●●●●●●●●●●●●●●
●●●●
●●
●●●●●●●●●●●●●
●
●
●
●●●●●● ●
●
●●●●
●
●●
●●
●●
●●
●
●●●
●
●
●●●●●●●●●●●●
●●●
●
●●
●
●
●
●
● ●
●●●●
●
● ●
●● ●●●
●●
●●
●●●●●●
●
●●
●● ●
●
●●●●
●●●
●●
●
●
●●●
●
●●●
●
●
●
●●
●●
●●●●●●●●●
●●
●
●●●●
●
●
●
●
●● ●●●●
●
●●●●
●●●
●
●●●●●●
●
●●●●●
●● ●●
●
●●
●● ●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●●
●
●●●
●
●●
●●
●●
●
●●
●
●●●
●
●●●●●●●●
●●●●●
●●●●●●●
●
●
●●
●● ●
● ●●●
●●
●●
●
●
●
●●
●
●●●
●●●●●
●●●●●
●●
●
● ●●●
●●●
●●●●●●
●
●
●
●●●
●●
●●
●●
●●
●●
●●●●●●
●●●●
●●●●●●●●●●●
●
●
●●
●●
●
●
●
●
●
●
AttainmentNonattainment
EPA regulatesall locations
●●
●
●
●
●●●●
●●●●●●●●●●●●●●
●●●●
●●
●●●●●●●●●●●●●
●
●
●
●●●●●● ●
●
●●●●
●
●●
●●
●●
●●
●
●●●
●
●
●●●●●●●●●●●●
●●●
●
●●
●
●
●
●
● ●
●●●●
●
● ●
●● ●●●
●●
●●
●●●●●●
●
●●
●● ●
●
●●●●
●●●
●●
●
●
●●●
●
●●●
●
●
●
●●
●●
●●●●●●●●●
●●
●
●●●●
●
●
●
●
●● ●●●●
●
●●●●
●●●
●
●●●●●●
●
●●●●●
●● ●●
●
●●
●● ●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●●
●
●●●
●
●●
●●
●●
●
●●
●
●●●
●
●●●●●●●●
●●●●●
●●●●●●●
●
●
●●
●● ●
● ●●●
●●
●●
●
●
●
●●
●
●●●
●●●●●
●●●●●
●●
●
● ●●●
●●●
●●●●●●
●
●
●
●●●
●●
●●
●●
●●
●●
●●●●●●
●●●●
●●●●●●●●●●●
●
●
●●
●●
●
●
●
●
●
●
AttainmentNonattainment
No EPA Action
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 5 /21
Observed Regulation Programand Overall Causal Effect
Observed Regulation Program
●●
●
●
●
●●●●
●●●●●●●●●●●●●●
●●●●
●●
●●●●●●●●●●●●●
●
●
●
●●●●●● ●
●
●●●●
●
●●
●●
●●
●●
●
●●●
●
●
●●●●●●●●●●●●
●●●
●
●●
●
●
●
●
● ●
●●●●
●
● ●
●● ●●●
●●
●●
●●●●●●
●
●●
●● ●
●
●●●●
●●●
●●
●
●
●●●
●
●●●
●
●
●
●●
●●
●●●●●●●●●
●●
●
●●●●
●
●
●
●
●● ●●●●
●
●●●●
●●●
●
●●●●●●
●
●●●●●
●● ●●
●
●●
●● ●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●●
●
●●●
●
●●
●●
●●
●
●●
●
●●●
●
●●●●●●●●
●●●●●
●●●●●●●
●
●
●●
●● ●
● ●●●
●●
●●
●
●
●
●●
●
●●●
●●●●●
●●●●●
●●
●
● ●●●
●●●
●●●●●●
●
●
●
●●●
●●
●●
●●
●●
●●
●●●●●●
●●●●
●●●●●●●●●●●
●
●
●●
●●
●
●
●
●
●
●
AttainmentNonattainment
Observational unit:pollution monitor(point-referenced data)
Overall Causal Effect
●●
●
●
●
●●●●
●●●●●●●●●●●●●●
●●●●
●●
●●●●●●●●●●●●●
●
●
●
●●●●●● ●
●
●●●●
●
●●
●●
●●
●●
●
●●●
●
●
●●●●●●●●●●●●
●●●
●
●●
●
●
●
●
● ●
●●●●
●
● ●
●● ●●●
●●
●●
●●●●●●
●
●●
●● ●
●
●●●●
●●●
●●
●
●
●●●
●
●●●
●
●
●
●●
●●
●●●●●●●●●
●●
●
●●●●
●
●
●
●
●● ●●●●
●
●●●●
●●●
●
●●●●●●
●
●●●●●
●● ●●
●
●●
●● ●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●●
●
●●●
●
●●
●●
●●
●
●●
●
●●●
●
●●●●●●●●
●●●●●
●●●●●●●
●
●
●●
●● ●
● ●●●
●●
●●
●
●
●
●●
●
●●●
●●●●●
●●●●●
●●
●
● ●●●
●●●
●●●●●●
●
●
●
●●●
●●
●●
●●
●●
●●
●●●●●●
●●●●
●●●●●●●●●●●
●
●
●●
●●
●
●
●
●
●
●
AttainmentNonattainment
EPA regulatesall locations
●●
●
●
●
●●●●
●●●●●●●●●●●●●●
●●●●
●●
●●●●●●●●●●●●●
●
●
●
●●●●●● ●
●
●●●●
●
●●
●●
●●
●●
●
●●●
●
●
●●●●●●●●●●●●
●●●
●
●●
●
●
●
●
● ●
●●●●
●
● ●
●● ●●●
●●
●●
●●●●●●
●
●●
●● ●
●
●●●●
●●●
●●
●
●
●●●
●
●●●
●
●
●
●●
●●
●●●●●●●●●
●●
●
●●●●
●
●
●
●
●● ●●●●
●
●●●●
●●●
●
●●●●●●
●
●●●●●
●● ●●
●
●●
●● ●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●●
●
●●●
●
●●
●●
●●
●
●●
●
●●●
●
●●●●●●●●
●●●●●
●●●●●●●
●
●
●●
●● ●
● ●●●
●●
●●
●
●
●
●●
●
●●●
●●●●●
●●●●●
●●
●
● ●●●
●●●
●●●●●●
●
●
●
●●●
●●
●●
●●
●●
●●
●●●●●●
●●●●
●●●●●●●●●●●
●
●
●●
●●
●
●
●
●
●
●
AttainmentNonattainment
No EPA Action
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 5 /21
Observed Regulation Programand Overall Causal Effect
Observed Regulation Program
●●
●
●
●
●●●●
●●●●●●●●●●●●●●
●●●●
●●
●●●●●●●●●●●●●
●
●
●
●●●●●● ●
●
●●●●
●
●●
●●
●●
●●
●
●●●
●
●
●●●●●●●●●●●●
●●●
●
●●
●
●
●
●
● ●
●●●●
●
● ●
●● ●●●
●●
●●
●●●●●●
●
●●
●● ●
●
●●●●
●●●
●●
●
●
●●●
●
●●●
●
●
●
●●
●●
●●●●●●●●●
●●
●
●●●●
●
●
●
●
●● ●●●●
●
●●●●
●●●
●
●●●●●●
●
●●●●●
●● ●●
●
●●
●● ●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●●
●
●●●
●
●●
●●
●●
●
●●
●
●●●
●
●●●●●●●●
●●●●●
●●●●●●●
●
●
●●
●● ●
● ●●●
●●
●●
●
●
●
●●
●
●●●
●●●●●
●●●●●
●●
●
● ●●●
●●●
●●●●●●
●
●
●
●●●
●●
●●
●●
●●
●●
●●●●●●
●●●●
●●●●●●●●●●●
●
●
●●
●●
●
●
●
●
●
●
AttainmentNonattainment
Observational unit:pollution monitor(point-referenced data)
Overall Causal Effect
●●
●
●
●
●●●●
●●●●●●●●●●●●●●
●●●●
●●
●●●●●●●●●●●●●
●
●
●
●●●●●● ●
●
●●●●
●
●●
●●
●●
●●
●
●●●
●
●
●●●●●●●●●●●●
●●●
●
●●
●
●
●
●
● ●
●●●●
●
● ●
●● ●●●
●●
●●
●●●●●●
●
●●
●● ●
●
●●●●
●●●
●●
●
●
●●●
●
●●●
●
●
●
●●
●●
●●●●●●●●●
●●
●
●●●●
●
●
●
●
●● ●●●●
●
●●●●
●●●
●
●●●●●●
●
●●●●●
●● ●●
●
●●
●● ●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●●
●
●●●
●
●●
●●
●●
●
●●
●
●●●
●
●●●●●●●●
●●●●●
●●●●●●●
●
●
●●
●● ●
● ●●●
●●
●●
●
●
●
●●
●
●●●
●●●●●
●●●●●
●●
●
● ●●●
●●●
●●●●●●
●
●
●
●●●
●●
●●
●●
●●
●●
●●●●●●
●●●●
●●●●●●●●●●●
●
●
●●
●●
●
●
●
●
●
●
AttainmentNonattainment
EPA regulatesall locations
●●
●
●
●
●●●●
●●●●●●●●●●●●●●
●●●●
●●
●●●●●●●●●●●●●
●
●
●
●●●●●● ●
●
●●●●
●
●●
●●
●●
●●
●
●●●
●
●
●●●●●●●●●●●●
●●●
●
●●
●
●
●
●
● ●
●●●●
●
● ●
●● ●●●
●●
●●
●●●●●●
●
●●
●● ●
●
●●●●
●●●
●●
●
●
●●●
●
●●●
●
●
●
●●
●●
●●●●●●●●●
●●
●
●●●●
●
●
●
●
●● ●●●●
●
●●●●
●●●
●
●●●●●●
●
●●●●●
●● ●●
●
●●
●● ●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●●
●
●●●
●
●●
●●
●●
●
●●
●
●●●
●
●●●●●●●●
●●●●●
●●●●●●●
●
●
●●
●● ●
● ●●●
●●
●●
●
●
●
●●
●
●●●
●●●●●
●●●●●
●●
●
● ●●●
●●●
●●●●●●
●
●
●
●●●
●●
●●
●●
●●
●●
●●●●●●
●●●●
●●●●●●●●●●●
●
●
●●
●●
●
●
●
●
●
●
AttainmentNonattainment
No EPA Action
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 6 /21
Potential Outcomes
Regulation program vector: A = [A(si)]ni=1
• n = 362 locations.• A(si) = 1⇒ i th location nonattainment.• Specific regulation program A = a.
Potential Outcomes• Pollution Xa(s): vector of pollutant concentrations.
• PM10 and O3 .• Mortality Ya(s): all-cause mortality among Medicare
beneficiaries living near a monitor.• ≈ 7 million people aged 65+.
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 6 /21
Potential Outcomes
Regulation program vector: A = [A(si)]ni=1
• n = 362 locations.• A(si) = 1⇒ i th location nonattainment.• Specific regulation program A = a.
Potential Outcomes• Pollution Xa(s): vector of pollutant concentrations.
• PM10 and O3 .• Mortality Ya(s): all-cause mortality among Medicare
beneficiaries living near a monitor.• ≈ 7 million people aged 65+.
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 7 /21
Spatial Correlation→Interference
• Typical assumption of no interference (SUTVA) likelyviolated.
• Regulations can affect air quality in other areas.
• Full interference: every location interferes with everyother location.
• 2n potential outcomes for each location.• Partial interference.
• Some locations interfere with some other locations.• How to define the interference groups?
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 7 /21
Spatial Correlation→Interference
• Typical assumption of no interference (SUTVA) likelyviolated.
• Regulations can affect air quality in other areas.• Full interference: every location interferes with every
other location.• 2n potential outcomes for each location.
• Partial interference.• Some locations interfere with some other locations.• How to define the interference groups?
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 7 /21
Spatial Correlation→Interference
• Typical assumption of no interference (SUTVA) likelyviolated.
• Regulations can affect air quality in other areas.• Full interference: every location interferes with every
other location.• 2n potential outcomes for each location.
• Partial interference.• Some locations interfere with some other locations.• How to define the interference groups?
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 8 /21
Never Observe PotentialOutcomes Under Regulation
Programs of Interest
Observed Regulation No Regulation Full RegulationA = aobs A = 0 A = 1
i Aobs(si ) Xaobs (si ) Yaobs (si ) X0(si ) Y0(si ) X1(si ) Y1(si )
Observed 1 0 obs obs
obs∗
obs∗ ? ?
Attainment 2 0 obs obs
obs∗ obs∗ ? ?
(black dots)...
.
.
....
.
.
....
.
.
....
.
.
.
Observed...
.
.
....
.
.
....
.
.
....
.
.
.nonattainment n − 1 1 obs obs
? ? obs∗ obs∗
(red dots) n 1 obs obs
? ? obs∗ obs∗
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 8 /21
Never Observe PotentialOutcomes Under Regulation
Programs of Interest
Observed Regulation No Regulation Full RegulationA = aobs A = 0 A = 1
i Aobs(si ) Xaobs (si ) Yaobs (si ) X0(si ) Y0(si ) X1(si ) Y1(si )
Observed 1 0 obs obs obs∗ obs∗ ? ?Attainment 2 0 obs obs obs∗ obs∗ ? ?
(black dots)...
.
.
....
.
.
....
.
.
....
.
.
.
Observed...
.
.
....
.
.
....
.
.
....
.
.
.nonattainment n − 1 1 obs obs ? ? obs∗ obs∗
(red dots) n 1 obs obs ? ? obs∗ obs∗
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 9 /21
Assignment Group InterferenceAssumption (AGIA)
●●
●
●
●
●●●●
●●●●●●●●●●●●●●
●●●●
●●
●●●●●●●●●●●●●
●
●
●
●●●●●● ●
●
●●●●
●
●●
●●
●●
●●
●
●●●
●
●
●●●●●●●●●●●●
●●●
●
●●
●
●
●
●
● ●
●●●●
●
● ●
●● ●●●
●●
●●
●●●●●●
●
●●
●● ●
●
●●●●
●●●
●●
●
●
●●●
●
●●●
●
●
●
●●
●●
●●●●●●●●●
●●
●
●●●●
●
●
●
●
●● ●●●●
●
●●●●
●●●
●
●●●●●●
●
●●●●●
●● ●●
●
●●
●● ●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●●
●
●●●
●
●●
●●
●●
●
●●
●
●●●
●
●●●●●●●●
●●●●●
●●●●●●●
●
●
●●
●● ●
● ●●●
●●
●●
●
●
●
●●
●
●●●
●●●●●
●●●●●
●●
●
● ●●●
●●●
●●●●●●
●
●
●
●●●
●●
●●
●●
●●
●●
●●●●●●
●●●●
●●●●●●●●●●●
●
●
●●
●●
●
●
●
●
●
●
AttainmentNonattainment
Interference implicitin EPA decisions
⇓Black dots don’tinterfere withRed dots
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 10 /21
Estimation
Models• Potential pollution outcomes.
• Multivariate spatial hierarchical model.• Mortality outcomes, conditional on pollution.
• Poisson regression.
• Estimation with data augmentation / MCMC.
Average Causal Effects
• Expected K-Dissociative Effect:• Average effect on mortality in areas where regulation
did not affect pollution.• Expected K-Associative Effect:
• Average effect on mortality in areas where regulationdecreased pollution.
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 10 /21
Estimation
Models• Potential pollution outcomes.
• Multivariate spatial hierarchical model.• Mortality outcomes, conditional on pollution.
• Poisson regression.
• Estimation with data augmentation / MCMC.
Average Causal Effects
• Expected K-Dissociative Effect:• Average effect on mortality in areas where regulation
did not affect pollution.
• Expected K-Associative Effect:• Average effect on mortality in areas where regulation
decreased pollution.
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 10 /21
Estimation
Models• Potential pollution outcomes.
• Multivariate spatial hierarchical model.• Mortality outcomes, conditional on pollution.
• Poisson regression.
• Estimation with data augmentation / MCMC.
Average Causal Effects
• Expected K-Dissociative Effect:• Average effect on mortality in areas where regulation
did not affect pollution.• Expected K-Associative Effect:
• Average effect on mortality in areas where regulationdecreased pollution.
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 11 /21
Spatial Hierarchical Model
X (s) = Z T (s)β + W (s) + ε(s)
• s ≡ specific location.
• X (s) ≡ 4−dimensional vector of pollution concentrations underboth regulations (XA=0(s),XA=1(s)).
• Z (s) ≡ covariates.
• ε(s) ≡ nonspatial (”nugget”) error.
• W (s) ≡ spatially-varying random intercepts.
• W (s) ∼ Multivariate Gaussian Process (MVGP).• Cross-covariance: K (s, s′; ν).
• ν governs spatial decay and smoothness.
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 11 /21
Spatial Hierarchical Model
X (s) = Z T (s)β + W (s) + ε(s)
• s ≡ specific location.
• X (s) ≡ 4−dimensional vector of pollution concentrations underboth regulations (XA=0(s),XA=1(s)).
• Z (s) ≡ covariates.
• ε(s) ≡ nonspatial (”nugget”) error.
• W (s) ≡ spatially-varying random intercepts.
• W (s) ∼ Multivariate Gaussian Process (MVGP).• Cross-covariance: K (s, s′; ν).
• ν governs spatial decay and smoothness.
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 12 /21
K (s, s′; ν)
Not RegulatedRegulated
sPM
O3
PM
O3
s'PM
O3
PM
O3
K (s, s; ν)
?
K (s′, s′; ν)
?K (s, s; ν) =K (s′, s′; ν)⇒ stationary
K (s, s′; ν) =( · · · ·· · · ·· · · ·· · · ·
)@@@R
��������n
n@@@R
��������
K (s, s′; ν) =( • · · ·· · · ·· · · ·· · · ·
)Corr(PM(s),PM(s′)) = e−ν||s−s′|| (isotropic)
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 12 /21
K (s, s′; ν)
Not RegulatedRegulated
sPM
O3
PM
O3
s'PM
O3
PM
O3K (s, s; ν)
?
K (s′, s′; ν)
?K (s, s; ν) =K (s′, s′; ν)⇒ stationary
K (s, s′; ν) =( · · · ·· · · ·· · · ·· · · ·
)@@@R
��������n
n@@@R
��������
K (s, s′; ν) =( • · · ·· · · ·· · · ·· · · ·
)Corr(PM(s),PM(s′)) = e−ν||s−s′|| (isotropic)
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 12 /21
K (s, s′; ν)
Not RegulatedRegulated
sPM
O3
PM
O3
s'PM
O3
PM
O3
K (s, s; ν)
?
K (s′, s′; ν)
?K (s, s; ν) =K (s′, s′; ν)⇒ stationary
K (s, s′; ν) =( · · · ·· · · ·· · · ·· · · ·
)@@@R
��������
nn@
@@R
��������
K (s, s′; ν) =( • · · ·· · · ·· · · ·· · · ·
)Corr(PM(s),PM(s′)) = e−ν||s−s′|| (isotropic)
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 12 /21
K (s, s′; ν)
Not RegulatedRegulated
sPM
O3
PM
O3
s'PM
O3
PM
O3
K (s, s; ν)
?
K (s′, s′; ν)
?K (s, s; ν) =K (s′, s′; ν)⇒ stationary
K (s, s′; ν) =( · · · ·· · · ·· · · ·· · · ·
)@@@R
��������
nn@
@@R
��������
K (s, s′; ν) =( • · · ·· · · ·· · · ·· · · ·
)
Corr(PM(s),PM(s′)) = e−ν||s−s′|| (isotropic)
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 12 /21
K (s, s′; ν)
Not RegulatedRegulated
sPM
O3
PM
O3
s'PM
O3
PM
O3
K (s, s; ν)
?
K (s′, s′; ν)
?K (s, s; ν) =K (s′, s′; ν)⇒ stationary
K (s, s′; ν) =( · · · ·· · · ·· · · ·· · · ·
)@@@R
��������
nn@
@@R
��������
K (s, s′; ν) =( • · · ·· · · ·· · · ·· · · ·
)
Corr(PM(s),PM(s′)) = e−ν||s−s′|| (isotropic)
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 13 /21
Specifying Spatial Structure
1 Specify K (s, s; ν)• Covariance matrix for potential pollution concentrations
within a location.• Nonidentifiability⇒ sensitivity parameter.
2 Specify spatial decay for each individual pollutionconcentration.
• Separate isotropic exponential decay functions for eachpollutant.
3 Combine 1. and 2. ⇒ cross-covariance function forMVGP.
• Computational feasibility.• Isoloate nonidentifiable associations between potential
outcomes.
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 14 /21
Why a Spatial Model?
Predicting potential outcomes:
• Predict missing potential outcomes at s usinginformation at surrounding locations.
• Use estimates of ν to assess interference assumption.
Not RegulatedRegulated
●
●●
●●
●●
● ●●
●●●●●●
●
●●
●●
●●
●
●
●
●
●
●●
●●● ●
●
●●
●
●
●●
●
●
●●
●
●
●
●●●
●
●●●●
●
●● ●●
●●
●
●
●●
●
●
●
●
●
● ●
●●●●
●
●● ●
●●
●
●
●
●
●●●
●●
●
●
●
●
●
● ●
●
●
●
●●
●●
●
●●
●
●
●●●
●
●●
●
●
●
●
●
●
●●
●●
●●●●●
●●
●●
●●
●
●
●
●●●●●●
●
●●
●
●●●
●
●●●●
●
●●●
●
●
s
iν
����
ν &%'$
ν
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 14 /21
Why a Spatial Model?
Predicting potential outcomes:
• Predict missing potential outcomes at s usinginformation at surrounding locations.
• Use estimates of ν to assess interference assumption.
Not RegulatedRegulated
●
●●
●●
●●
● ●●
●●●●●●
●
●●
●●
●●
●
●
●
●
●
●●
●●● ●
●
●●
●
●
●●
●
●
●●
●
●
●
●●●
●
●●●●
●
●● ●●
●●
●
●
●●
●
●
●
●
●
● ●
●●●●
●
●● ●
●●
●
●
●
●
●●●
●●
●
●
●
●
●
● ●
●
●
●
●●
●●
●
●●
●
●
●●●
●
●●
●
●
●
●
●
●
●●
●●
●●●●●
●●
●●
●●
●
●
●
●●●●●●
●
●●
●
●●●
●
●●●●
●
●●●
●
●
s
iν
����
ν
&%'$
ν
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 14 /21
Why a Spatial Model?
Predicting potential outcomes:
• Predict missing potential outcomes at s usinginformation at surrounding locations.
• Use estimates of ν to assess interference assumption.
Not RegulatedRegulated
●
●●
●●
●●
● ●●
●●●●●●
●
●●
●●
●●
●
●
●
●
●
●●
●●● ●
●
●●
●
●
●●
●
●
●●
●
●
●
●●●
●
●●●●
●
●● ●●
●●
●
●
●●
●
●
●
●
●
● ●
●●●●
●
●● ●
●●
●
●
●
●
●●●
●●
●
●
●
●
●
● ●
●
●
●
●●
●●
●
●●
●
●
●●●
●
●●
●
●
●
●
●
●
●●
●●
●●●●●
●●
●●
●●
●
●
●
●●●●●●
●
●●
●
●●●
●
●●●●
●
●●●
●
●
s
iν ����
ν
&%'$
ν
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 15 /21
Assess Interference Assumption
ν̂ has implications of interference
• ν̂ ⇒ estimated correlation between measurements attwo locations.
• Examine correlations between observations assumednot to interfere.
• Substantial correlation⇒ violation of AGIA.
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 16 /21
Assessment of AGIA for PM10
Figure: PM10 , ν̂ = 3.13
Correlation at distance ds
Freq
uenc
y
0.0 0.1 0.2 0.3 0.4 0.5 0.6
020
4060
8010
0
Correlation > 0.2530 locations
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 17 /21
Assessment of AGIA for O3
Figure: O3 , ν̂ = 2.68
Correlation at distance ds
Freq
uenc
y
0.0 0.1 0.2 0.3 0.4 0.5 0.6
010
2030
4050
60
Correlation > 0.2554 locations
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 18 /21
Associative and DissociativeEffects for PM10
●
●
●
●
●
●
●●
●●
●
●
●●●
●●●
●
●
●
●
●
●
●
●●●●
●
●●
●
●●●
●
●
●
●●●
●
●
●
●
●
●●
●
●●
Dea
ths/
1000
EDE EAE
−8−6
−4−2
02
4
ω=0 ω=0.3 ω=0.6 ω=0.9 ω=0 ω=0.3 ω=0.6 ω=0.9
Causal effect on heath in areas where:regulation does not
affect PM10regulation causally
reduces PM10
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 19 /21
Associative and DissociativeEffects for joint effect on both
PM10 and O3
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●●
●●●
●●
●
●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●●
●●
●
●●●●
●
●●●
●
●●●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●
● ●●
●
●
●
●
●●●
Dea
ths/
1000
EDE EAE−1
0−8
−6−4
−20
24
68
ω=0 ω=0.3 ω=0.6 ω=0.9 ω=0 ω=0.3 ω=0.6 ω=0.9
Causal effect on heath in areas where:regulation does not
affect PM10 or O3regulation causally
reduces PM10 and O3
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 20 /21
Summary
Causal inference for accountability assessment
• Complex regulatory environment.• Causal inference with spatial data.• Principal stratification
• Multivariate intermediate variable.• Multipollutant approach.
• Assumptions about interference between observations.• What do we assume?• How do we assess?• What are the implications of violations?
CausalInference forAir Quality
Regulations
Corwin M.Zigler
Accountabilityfor Air Quality
PotentialOutcomesandInterference
SpatialHierarchicalModel
AssessingInterferenceAssumptions
Analysis ofthe 1990Clean Air ActAmendments
[email protected] 21 /21
Thank You
Acknowledgments
• Francesca Dominici• Yun Wang• Funding: Health Effects Institute.• Publication: Zigler CM, Dominici F, and Wang Y.
Estimating causal effects of air quality regulations usingprincipal stratification for spatially-correlatedmultivariate intermediate outcomes. Biostatistics 2012;13(2): 289–302.