Motivation
•
WHO estimated that ambient air pollution occasions annual health cost of US$
1.27 trillion in Europe (WHO Regional Office for Europe and OECD, 2015)
•
Precise dose-response estimates needed that relate observable health outcomes to
measures of local pollution exposure
Challenges to identify the dose-response relationship
•
People living in areas with poor air quality may have a worse health status for
reasons that are unrelated to air pollution exposure
•
People may decide to live in areas with good air quality because they derive utility
from unobserved location characteristics confounded with air quality
Overcoming the identification challenge
•
Economic literature has suggested to use fixed effects approach to address the
unobserved heterogeneity issue
•
Link between recession, air pollution and infant mortality (Chay and Greenstone,
2003)
•
Ambient air pollution, infant birth outcomes and infant mortality (Currie et al.,
2009)
•
Traffic emission and infant mortality (Knittel et al., 2015)
•
Insufficient knowledge on the relationship between ambient air pollution and
morbidity
•
Hospital admission and air pollution (Lagravinese et al., 2014)
Endogenous treatment variable?
•
Standard approach assumes that air pollutant concentration is homogenous within
a given region −→ potential of measurement bias in the treatment variable
•
Alternative approach relies on spatial interpolation (less strict with respect to the
homogeneity assumption)
•
Downside of the weighting approach is that it does not account for emission sources
and atmospheric conditions
•
Economic literature has resorted to instrumental variable estimation techniques to
address the endogeneity problem
•
IV approach only applicable when appropriate instruments are available
Alternative solution
•
Solve the measurement issue at the source and construct local measures of air
pollution with a dispersion model that replicates the atmospheric conditions while
accounting for emission sources
Why the interpolation method matters?
(a) Station data interpolated
(b) Dispersion model interpolated
Average NO2 pollution exposure for 2001/2012
Empirical strategy
•
Outcome variable is hospital visits (non-negative and discrete nature) −→
pseudo-maximum likelihood Poisson model (Gourieroux et al., 1984; Colin Cameron
and Trivedi, 2013)
adm
it= exp (α
i+ p
itβ + X
itγ
x+ δ
ct) ρ
it(1)
•
Location fixed effects denoted by α
iand canton-time fixed effects by δ
ct•
Treatment variable is p
itmeasuring average pollution exposure for PM10, NO2,
SO2 and O3
•
X
itis a matrix of covariates (count of hospital admissions for physiological
diseases, income, equality, population, foreigner share, unemployment rate)
•
Accounting for clustering at the region level with robust variance estimator (Colin
Investigated causes of hospital admissions
Cause of hospital admissions ICD-10 code Description
All cardiovascular diseases I00-I99 All diseases that are related to the cardiovascular system.
Coronary artery disease I20-I25 Stable angina, unstable angina, myocardial infarc-tion, and sudden coronary death.
Cerebrovascular disease I60-I69 Vascular disease of the cerebral circulation. All respiratory diseases J00-J99 All conditions of the upper respiratory tract,
tra-chea, bronchi, bronchioles, alveoli, pleura and pleural cavity, and the nerves and muscles of breathing.
Pneumonia J12-J18, P23 Inflammatory condition of the lung.
COPD J40-J44 Obstructive lung disease characterized by chroni-cally poor airflow.
Asthma J45-J46 Chronic inflammatory disease characterized by variable and recurring symptoms, reversible airflow obstruction and bronchospasm.
Diabetes E10-E14 Metabolic disease in which there are high blood sugar levels over a prolonged period.
From dispersion output to local pollution exposure
Dispersion
model output
Spatial selection
of settlement area
Spatial assignment at
MedStat region level
Spatial variation in pollution exposure
From station data to local pollution exposure
•
Air pollution data for >130 monitoring
stations from the Swiss National Air
Pollution Monitoring Network (NABEL)
•
Interpolate pollution exposure at the
MedStat region level with ‘naive’
inverse-distance weighting approach
b
p
it=
nX
j=11
d
ijp
jt/
nX
j=11
d
ij•
Calculate annual pollution exposure for
Spatial variation in pollution exposure
Ambient air pollution and hospital admissions (dispersion model approach)
PM10 NO2 SO2 O3 All cardiovascular diseases 1.331 2.736∗∗∗ 4.910∗∗∗ 0.132
(0.823) (1.002) (1.403) (0.127) Coronary artery disease 1.292 4.385∗∗∗ 4.133∗∗∗ 0.207
(1.168) (1.344) (1.555) (0.195) Cerebrovascular disease −0.219 2.564∗ 5.844∗∗∗ −0.079
(1.402) (1.476) (1.978) (0.258) All respiratory diseases −0.577 0.346 3.408∗∗ −0.064
(0.966) (1.039) (1.388) (0.150) Pneumonia −1.233 −0.348 1.586 −0.130 (1.481) (1.562) (1.710) (0.240) COPD −3.134 −0.887 4.033∗ −0.264 (2.266) (2.046) (2.388) (0.379) Asthma 4.715 −1.326 5.309 0.188 (3.604) (3.100) (3.275) (0.541) Diabetes −2.549 0.527 3.286 −0.496 (2.101) (1.867) (2.210) (0.397) Diseases of middle ear and mastoid −3.611 0.319 1.076 0.366
Ambient air pollution and hospital admissions (station data approach)
PM10 NO2 SO2 O3 All cardiovascular diseases 0.123 0.054 −0.466 0.073 (0.292) (0.153) (0.288) (0.067) Coronary artery disease 0.149 −0.205 −1.567∗∗∗ 0.106
(0.380) (0.217) (0.375) (0.090) Cerebrovascular disease 0.047 0.100 −0.933∗ −0.086
(0.473) (0.275) (0.477) (0.107) All respiratory diseases −0.071 0.264∗ −0.600∗ −0.124
(0.307) (0.159) (0.332) (0.076) Pneumonia 0.237 0.122 −1.579∗∗∗ −0.078 (0.446) (0.229) (0.490) (0.121) COPD −0.036 0.627∗ −0.480 −0.165 (0.616) (0.363) (0.718) (0.172) Asthma 0.777 −1.033 −1.174 0.175 (1.029) (0.640) (1.327) (0.251) Diabetes −0.688 −0.042 1.255 −0.155 (0.585) (0.342) (0.769) (0.228) Diseases of middle ear and mastoid −2.529∗∗ −1.260∗∗ −2.056∗ 0.254
Testing for non-linearity in the treatment effect
•
Epidemiological literature suggests that the relationship between air pollution and
hospital admission is non-linear in the level of pollution exposure (e.g., Pope et al.,
2009)
•
Question: Does high pollution exposure matters more for predicting negative health
outcomes than moderate pollution exposure?
•
Schlenker & Walker (2015) provide evidence that the dose-response function is not
convex (decreasing marginal effect of pollution exposure)
•
Test for non-linearity in the treatment effect with quantile specification of the
Non-linearity in the NO2 treatment effect
Q1 Q2 Q3 Q4
All cardiovascular diseases 2.463∗∗ 2.512∗∗ 2.590∗∗ 2.622∗∗
(1.084) (1.053) (1.031) (1.020) Coronary artery disease 4.145∗∗∗ 4.229∗∗∗ 4.442∗∗∗ 4.321∗∗∗
(1.430) (1.397) (1.370) (1.358) Cerebrovascular disease 2.233 2.237 2.126 2.369
(1.604) (1.530) (1.500) (1.484) All respiratory diseases 0.279 0.311 0.355 0.327
(1.132) (1.087) (1.064) (1.054) Pneumonia −0.973 −0.601 −0.533 −0.556 (1.691) (1.640) (1.607) (1.586) COPD −0.959 −1.462 −1.235 −1.033 (2.342) (2.190) (2.146) (2.099) Asthma −0.923 −1.694 −1.841 −1.354 (3.407) (3.295) (3.227) (3.158) Diabetes 0.611 0.359 0.248 0.478 (2.135) (2.062) (2.013) (1.931) Disesaes of middle ear and mastoid 1.206 0.533 0.821 0.624
Non-linearity in the SO2 treatment effect
Q1 Q2 Q3 Q4
All cardiovascular diseases 4.178∗ 2.780 4.226∗∗∗ 4.422∗∗∗
(2.447) (1.770) (1.557) (1.385) Coronary artery disease 4.092 2.560 4.167∗∗ 3.906∗∗
(2.993) (2.145) (1.806) (1.579) Cerebrovascular disease −0.128 −0.435 3.405 4.095∗∗
(3.845) (2.771) (2.293) (1.932) All respiratory diseases 3.668 1.332 3.217∗∗ 3.073∗∗
(2.290) (1.771) (1.568) (1.363) Pneumonia −3.062 −3.563 0.221 0.280 (3.113) (2.350) (1.986) (1.679) COPD 9.809∗∗ 2.214 3.788 4.061∗ (4.233) (3.452) (2.866) (2.459) Asthma 14.404∗∗ 12.345∗∗∗ 10.411∗∗∗ 7.805∗∗ (6.944) (4.518) (3.944) (3.417) Diabetes 2.917 0.370 2.163 2.624 (4.964) (3.126) (2.526) (2.261) Disesaes of middle ear and mastoid −9.991 −2.174 −0.924 −0.346
Heterogenous response to air pollution exposure
•
Strong claim in the epidemiological literature that ambient air pollution has a
different effect on on emergency admissions than on elective admissions (e.g.,
Perez et al., 2015)
•
Patients often not admitted overnight when they suffer from respiratory distress
(Schlenker & Walker, 2015)
•
Question 1: Does the treatment effect vary by admission type?
•
Heterogenous response to air pollution between age groups reported in the
literature
•
Question 2: Do elderly people react differently to pollution exposure than the
Variation in the treatment effect by admission type
Emergency admissions Elective admissions NO2 SO2 NO2 SO2 All cardiovascular diseases 1.553 5.311∗∗∗ 2.740∗∗ 3.908∗∗∗
(1.313) (1.671) (1.128) (1.468) Coronary artery disease 2.337 3.161∗ 5.329∗∗∗ 4.222∗∗
(1.722) (1.730) (1.574) (1.836) Cerebrovascular disease 1.174 6.688∗∗∗ 1.882 2.937
(1.650) (2.092) (2.281) (2.819) All respiratory diseases −0.677 5.290∗∗∗ −0.532 −0.390
(1.451) (1.932) (1.293) (1.300) Pneumonia −0.898 2.555 −8.114∗∗ −12.297∗∗∗ (1.750) (2.007) (3.852) (3.929) COPD −3.608 3.808 2.132 2.989 (2.198) (2.375) (3.390) (4.193) Asthma −1.130 4.511 −6.784 6.450 (3.303) (3.568) (6.249) (7.072) Diabetes 0.468 3.826 −1.376 0.164 (2.258) (2.674) (2.873) (2.778)
Age differences in the response to air pollution
Working age population Elderly population NO2 SO2 NO2 SO2 All cardiovascular diseases 3.093∗∗∗ 3.843∗∗∗ 2.836∗∗ 6.006∗∗∗
(1.076) (1.327) (1.191) (1.665) Coronary artery disease 5.217∗∗∗ 3.218∗ 3.983∗∗ 4.905∗∗∗
(1.661) (1.760) (1.610) (1.821) Cerebrovascular disease −0.176 2.676 3.516∗∗ 6.902∗∗∗
(2.263) (2.269) (1.687) (2.296) All respiratory diseases 1.573 3.664∗∗∗ −1.476 1.881
(1.155) (1.284) (1.498) (1.798) Pneumonia 0.600 3.882∗ −1.018 −0.106 (2.091) (2.015) (1.865) (1.810) COPD 3.878 5.293 −2.680 3.563 (3.543) (3.473) (2.270) (2.529) Asthma −0.548 4.836 −9.569∗ 0.775 (4.354) (4.320) (5.479) (4.853) Diabetes 0.326 2.335 0.050 3.830 (2.557) (2.470) (2.641) (3.267) Diseases of middle ear and mastoid 0.196 5.800∗ 33.423∗∗∗ 15.543∗ (4.161) (3.457) (9.919) (9.294)
Addressing endogeneity issues with the air pollution measure
•
Account for measurement error in the exposure measure applying control-function
approach to the conditional Poisson model (Terza et al., 2008; Colin Cameron and
Trivedi, 2013)
•
Step 1: Include spatial and temporal lags of air pollution measure in first stage
regression
p
it= α
i+ δ
ct+ p
i,t−1β
1+ ¯
p
itβ
2+ X
itγ
x+ u
it•
Step 2: Include first-stage residuals in the second stage regression of pollution
exposure on health outcomes
Estimation results for the control function approach
PM10 NO2 SO2 O3 All cardiovascular diseases 1.924∗ 4.249∗∗∗ 5.443∗∗∗ 0.117
(1.108) (1.396) (1.696) (0.144) Coronary artery disease 1.488 5.820∗∗∗ 3.376∗ 0.208
(1.534) (1.844) (1.930) (0.218) Cerebrovascular disease 0.401 3.296 5.495∗∗ −0.201
(1.843) (2.114) (2.170) (0.275) All respiratory diseases −1.165 0.678 3.694∗∗ −0.167
(1.211) (1.430) (1.735) (0.165) Pneumonia −1.968 −0.719 1.220 −0.278 (1.816) (2.076) (2.225) (0.257) COPD −3.498 −0.937 3.476 −0.237 (2.812) (2.776) (3.030) (0.426) Asthma 3.964 −4.510 3.758 0.365 (4.298) (4.192) (4.031) (0.592) Diabetes −1.963 1.034 2.676 −0.048 (3.050) (2.792) (2.691) (0.415) Disesaes of middle ear and mastoid 1.780 0.748 0.331 0.747
Alternative specification with spatiotemporal instrumental variables
PM10 NO2 SO2 O3 All cardiovascular diseases −0.859 8.208∗∗ 6.065∗∗∗ −0.010
(3.096) (3.908) (2.164) (0.826) Coronary artery disease −2.841 11.896∗∗ 3.167 −0.387
(4.171) (5.092) (2.575) (1.251) Cerebrovascular disease 1.006 2.651 5.817∗∗ −0.282
(4.643) (5.841) (2.815) (1.658) All respiratory diseases −2.504 −0.767 4.347∗∗ −2.216∗∗
(3.160) (4.179) (2.171) (1.014) Pneumonia −0.604 −8.065 1.935 −4.798∗∗∗ (5.411) (5.394) (2.911) (1.624) COPD −1.763 −0.258 2.727 −1.419 (7.529) (8.066) (4.148) (2.665) Asthma 14.271 −10.243 7.862 −5.577 (11.080) (11.423) (5.323) (4.144) Diabetes 0.166 8.892 2.633 −2.458 (8.625) (8.092) (3.562) (2.770) Disesaes of middle ear and mastoid −9.875 −2.254 0.588 −4.632