A fundamental feature of evaluating causal health effects of air quality regulations is that air pollution moves through space, rendering health outcomes at a particular population location dependent upon regulatory actions taken at multiple, possibly distant, pollution sources. Motivated by studies of the public-health impacts of power plant regulations in the U.S., this talk introduces the novel setting of bipartite causal inference with interference, which arises when 1) treatments are defined on observational units that are distinct from those at which outcomes are measured and 2) there is interference between units in the sense that outcomes for some units depend on the treatments assigned to many other units. Interference in this setting arises due to complex exposure patterns dictated by physical-chemical atmospheric processes of pollution transport, with intervention effects framed as propagating across a bipartite network of power plants and residential zip codes. New causal estimands are introduced for the bipartite setting, along with an estimation approach based on generalized propensity scores for treatments on a network. The new methods are deployed to estimate how emission-reduction technologies implemented at coal-fired power plants causally affect health outcomes among Medicare beneficiaries in the U.S.
Similar to Causal Inference Opening Workshop - Bipartite Causal Inference with Interference for Evaluating Air Pollution Regulations - Corwin Zigler, December 10, 2019
Similar to Causal Inference Opening Workshop - Bipartite Causal Inference with Interference for Evaluating Air Pollution Regulations - Corwin Zigler, December 10, 2019 (20)
Causal Inference Opening Workshop - Bipartite Causal Inference with Interference for Evaluating Air Pollution Regulations - Corwin Zigler, December 10, 2019
1. cory.zigler@austin.utexas.edu 1 /32
Bipartite Causal Inference with Interference
for Evaluating Air Pollution Regulations
Corwin M. Zigler
Department of Statistics and Data Sciences
Department of Women’s Health
University of Texas at Austin and Dell Medical School
December 10, 2019
3. cory.zigler@austin.utexas.edu 3 /32
Major Pollution Source: Power Plants
⇒ Many regulations to reduce emissions
• Some install “scrubbers” in
response to regulations
→ Reduce SO2 emissions
→ Reduce ambient
particulate matter (PM2.5 )
→ Improve health
∗Note: Due to regulations in the Clean Air
Act estimated to be the the most costly and
most beneficial of all monetized federal
regulations.
4. cory.zigler@austin.utexas.edu 3 /32
Major Pollution Source: Power Plants
⇒ Many regulations to reduce emissions
• Some install “scrubbers” in
response to regulations
→ Reduce SO2 emissions
→ Reduce ambient
particulate matter (PM2.5 )
→ Improve health
∗Note: Due to regulations in the Clean Air
Act estimated to be the the most costly and
most beneficial of all monetized federal
regulations.
Question: Do scrubbers on coal-fired power plants causally
affect hospitalizations for Ischemic Heart Disease (IHD) among
Medicare beneficiaries?
7. cory.zigler@austin.utexas.edu 6 /32
Interference Due to Treatment Diffusion
or Complex Exposure Dependencies
(compare with unit-to-unit outcome dependencies in a social network)
• Here, interference arises due to treatment diffusion
• Resulting in complex exposure dependencies
• Governed by atmospheric transport of pollution
• Scrubber impacts diffuse across a network
• Knowledge of the complex “interference process” is
available
• Deterministic physical-chemical models of air pollution
• Need some repurposing for this purpose
9. cory.zigler@austin.utexas.edu 8 /32
Defining Characteristics of Bipartite
Causal Inference with Interference
Evaluating the causal effects of a (binary) intervention where:
1. “Bipartite” Structure
• The intervention is defined on one type of observational
unit
• E.g., whether a coal-fired power plant installs a “scrubber”
• Outcomes of interest are defined and measured on a
second, distinct, type of unit
• E.g., hospitalizations measured at residential zip codes
10. cory.zigler@austin.utexas.edu 8 /32
Defining Characteristics of Bipartite
Causal Inference with Interference
Evaluating the causal effects of a (binary) intervention where:
1. “Bipartite” Structure
• The intervention is defined on one type of observational
unit
• E.g., whether a coal-fired power plant installs a “scrubber”
• Outcomes of interest are defined and measured on a
second, distinct, type of unit
• E.g., hospitalizations measured at residential zip codes
2. Interference
Interconnectedness between units ⇒ outcomes for a particular
unit depend upon treatments assigned to (multiple) other units
11. cory.zigler@austin.utexas.edu 9 /32
Formalization of Bipartite Structure
Two types of observational units:
1 Interventional Units
• P = {p1, p2, . . . , pi , . . . , pP}: a set of P power plants
• Where interventions occur (or not)
• Ai = 1/0 denotes presence/absence of intervention
• Treatment allocation: A = (A1, A2, . . . , AP)
• Covariates W = (W1, W2, . . . , WP)
12. cory.zigler@austin.utexas.edu 9 /32
Formalization of Bipartite Structure
Two types of observational units:
1 Interventional Units
• P = {p1, p2, . . . , pi , . . . , pP}: a set of P power plants
• Where interventions occur (or not)
• Ai = 1/0 denotes presence/absence of intervention
• Treatment allocation: A = (A1, A2, . . . , AP)
• Covariates W = (W1, W2, . . . , WP)
2 Outcome Units
• M = {m1, m2, . . . , mj , . . . , mM }: a set of M locations
• Where outcomes are measured/of interest
• Yj denotes outcome if interest
• Covariates: X = (X1, X2, . . . , XM )
13. cory.zigler@austin.utexas.edu 10 /32
Potential Outcomes with Bipartite
Interference
Yj (A) = Yj (A1 = a1, A2 = a2, . . . , AP = aP )
Ai : Treatment assignment of the the ith power plant
A(−i) Treatment assignments of all power plants except i
Most primitive individual-level causal effect:
Yj(Ai = a, A(−i) = a(−i)) − Yj(Ai = a , A(−i) = a(−i))
Effect on outcome unit mj of treatment allocation a with ai = a
versus treatment allocation a with ai = a
Key bipartite feature: P such effects defined for every
j = 1, 2, . . . , M
Complicates typical notions of which treatments “directly” apply
to a unit
14. cory.zigler@austin.utexas.edu 11 /32
Direct and Indirect (or Spillover) Effects
Direct Effect (at individual M-level)
DE(i,j) = Yj(Ai = 1, A(−i) = a) − Yj(Ai = 0, A(−i) = a)
• Effect on location j of intervening (vs. not) at power plant i,
under a given allocation to all other plants
15. cory.zigler@austin.utexas.edu 11 /32
Direct and Indirect (or Spillover) Effects
Direct Effect (at individual M-level)
DE(i,j) = Yj(Ai = 1, A(−i) = a) − Yj(Ai = 0, A(−i) = a)
• Effect on location j of intervening (vs. not) at power plant i,
under a given allocation to all other plants
Indirect Effect (at individual M-level)
IE(i,j) = Yj(Ai = 0, A(−i) = a) − Yj(Ai = 0, A(−i) = a )
• Effect of changing intervention allocation to other upwind
plants
16. cory.zigler@austin.utexas.edu 12 /32
Many Types of Average Effects
owing to the (i, j) indexing of individual-level estimands
1 P-indexed effects
• Effects of treatments on pi ∈ P, averaged over locations
mj ∈ M that contain pi in their interference set
• E.g., the “typical” effect across all connected locations of
treating a single power plant
17. cory.zigler@austin.utexas.edu 12 /32
Many Types of Average Effects
owing to the (i, j) indexing of individual-level estimands
1 P-indexed effects
• Effects of treatments on pi ∈ P, averaged over locations
mj ∈ M that contain pi in their interference set
• E.g., the “typical” effect across all connected locations of
treating a single power plant
2 M-indexed effects
• Effects on mj ∈ M, averaged over treatments at pi ∈ P in
the interference set
• E.g., the “typical” effect on a single location of treating a
single power plant
18. cory.zigler@austin.utexas.edu 12 /32
Many Types of Average Effects
owing to the (i, j) indexing of individual-level estimands
2 M-indexed effects
• Effects on mj ∈ M, averaged over treatments at pi ∈ P in
the interference set
• E.g., the “typical” effect on a single location of treating a
single power plant
→ Subset of M∗-indexed effects
• Do not average over all (i, j)
• Effects on mj ∈ M, defined according to a single i∗
in the
interference set
• E.g., the “typical” effect of treating a key associated i∗
19. cory.zigler@austin.utexas.edu 13 /32
“Key-Associated” Causal Effects
In the bipartite setting
In this talk, estimands will be developed based on the following
rationale (although more general development available):
• Consider for each mj, a particular power plant of interest
• E.g., the closest (or most influential) plant to location j
• Denote with p∗
i(j)
• Simplify notation: p∗
i(j) ≡ i∗
• For each j, only consider one
Yj(Ai∗ = a, A(−i∗) = a(−i∗)), DE(i∗,j), IE(i∗,j)
22. cory.zigler@austin.utexas.edu 15 /32
HYSPLIT Average Dispersions
(HyADS)
Computationally scalable approximation of air pollution transport
• Model movement of air
parcel through time/space
• Repeat multiple times per
power plant per hour per
day per year, . . .
• ≈ 108 parcels per year
tracked in time and space
23. cory.zigler@austin.utexas.edu 15 /32
HYSPLIT Average Dispersions
(HyADS)
Computationally scalable approximation of air pollution transport
• ⇒ tij to denote “influence”
of pi on mj
• tj = (tji, tj2, . . . , tjP) ≡
“interference map” for the
jth location
25. cory.zigler@austin.utexas.edu 17 /32
Bivariate Treatment
Forastiere et al. (2019+)
1 “Individual” Treatment: Scrubber status of the key
associated power plant: Ai∗
• i∗
denotes the power plant with the most influence on zip
code j
• i∗
≡ i; tji = maxi {tj }
• Ai∗ = 0, 1 according to whether the key power plant has a
scrubber
26. cory.zigler@austin.utexas.edu 17 /32
Bivariate Treatment
Forastiere et al. (2019+)
1 “Individual” Treatment: Scrubber status of the key
associated power plant: Ai∗
• i∗
denotes the power plant with the most influence on zip
code j
• i∗
≡ i; tji = maxi {tj }
• Ai∗ = 0, 1 according to whether the key power plant has a
scrubber
2 “Upwind” Treatment: Function of the scrubber statuses of
all other “linked” power plants
• “Exposure Mapping:” gj (·; T) : {0, 1}P−1
→ Gi
• gj (A, T) = Gj = i=i∗ tji Ai
• HyADS-weighted upwind treatment rate
27. cory.zigler@austin.utexas.edu 18 /32
“Upwind Interference” Assumption
(Adaptation of SUTVA)
For any A, A such that Ai∗ = Ai∗ and Gj = Gj :
Yj(A) = Yj(Ai∗ , Gj) = Yj(Ai∗ , Gj ) = Yj(A )
• Reduces interference to depend only on i∗ and Gj (not on
the entire A)
• Hospitalizations are the same under different A if the
treatment of the key power plant and the “upwind”
treatment rate are the same
• The key assumption about interference
28. cory.zigler@austin.utexas.edu 19 /32
“Direct” Effects in the Bipartite Setting
The “direct” effect of treating the key plant while holding fixed
the “upwind” treatments:
τ(g) = E Yj(Ai∗ = 1, Gj = g) − Yi(Ai∗ = 0, Gj = g)
“Overall” average over the distribution of “upwind” treatments:
τ =
g∈G
τ(g)P(Gj = g)
“The average effect on IHD hospitalizations of installing a
scrubber on the key associated power plant”
29. cory.zigler@austin.utexas.edu 20 /32
“Indirect” or “Spillover” Effects in the
Bipartite Setting
The “indirect” or “upwind” effect of installing scrubbers on
upwind power plants without changing the key plant:
δ(g; a) = E Yj(Ai∗ = a, Gj = g) − Yj(Ai∗ = a, Gj = 0)
“Overall” average over the distribution of “upwind” treatments:
∆(a) =
g∈G
δ(g; a)P(Gj = g)
“The average effect on IHD hospitalizations of installing more
scrubbers on upwind plants”
30. cory.zigler@austin.utexas.edu 21 /32
Key-Associated and Neighborhood
Propensity Scores
Forastiere et al. (2019+)
View “key-associated” and “upwind” treatments as bivariate
treatment and estimate a joint propensity score:
ψ(a; g; x) = P(Ai∗ = a, Gj = g|Xj = x)
Further decompose into:
ψ(a; g; x) = P(Ai∗ = a, Gj = g|Xj = x)
= P(Ai∗ = a|Xi∗ = x) (1)
× P(Gj = g|Ai∗ = a, X = x) (2)
(1) ≡ φ(a, x) ≡ “key-associated propensity sore”
(2) ≡ λ(g, z, x) ≡ “upwind propensity score”
31. cory.zigler@austin.utexas.edu 22 /32
Estimation Strategy (1/3)
Extension of Forastiere et al. (2019+), other variations on this theme
1. Estimate the key-associated propensity score
φ(a, x) = P(Ai∗ = a|Xj) = fA∗
(z, Xj; γ)
• Logistic regression
• Key-associated PS estimates: ˆφj for j = 1, 2, . . . ,27,009
2. Stratify zip codes
• Into K = 5 strata based on ˆφj
32. cory.zigler@austin.utexas.edu 23 /32
Estimation Strategy (2/3)
Extension of Forastiere et al. (2019+), other variations on this theme
Within each k = 1, 2, . . . , K strata:
3. Estimate the “upwind” propensity score
λ(g, z, x) = P(Gj = g|Ai∗ = a, Xj) = fG
(g, a, Xj; δk
)
• Normal regression model
• Upwind PS estimates: ˆλj
4. Estimate parametric dose-response function
Yj(a, g)|ˆλj ∼ fy
(a, g, ˆλ; θk
)
• Poisson regression model
• Predicted potential outcomes: ˆYj(a, g)
• For every (a, g) across a grid of g
33. cory.zigler@austin.utexas.edu 24 /32
Estimation Strategy (3/3)
Extension of Forastiere et al. (2019+), other variations on this theme
5. Calculate average causal estimates
• Straightforward with ˆYj(a, g) in hand
• “Direct Effect”: τ
• “Upwind Effects”: ∆(a) for a = 0, 1
6. Bootstrap for standard errors and intervals
• “Egocentric” bootstrap, independent samples of zip codes,
preserving network structure
• Not exactly right for this sampling mechanism, but shown
to be reasonable
35. cory.zigler@austin.utexas.edu 26 /32
Power Plant and Zip Code Data
Integrated with HyADS atmospheric model
472 Coal-fired power plants
operating in 2005
• 106 with scrubbers
installed
• Information on emissions,
plant size, operating
capacity, other controls,
etc.