Documente Academic
Documente Profesional
Documente Cultură
pubs.acs.org/est
Industrial and Systems Engineering Department, Ponticia Universidad Catolica de Chile, Santiago, Chile
S Supporting Information
*
ABSTRACT: To further the understanding and implementation of expert elicitation methods in the evaluation of public
policies related to air pollution, the present studys main goal
was to explore the potential strengths and weaknesses of
structured expert judgment (SEJ) methodology as a way to
derive a C-R function for chronic PM2.5 exposure and premature
mortality in Chile. Local experts were classied in two groups
according to background and experience: physicians (Group 1)
and engineers (Group 2). Experts were required to provide an
estimate of the true percent change in nonaccidental mortality
resulting from a permanent 1 g/m3 reduction in PM2.5 annual
average ambient concentration across the entire Chilean
territory. Cookes Classical Model was used to combine the
individual experts assessments. Experts mortality estimations varied markedly across groups: while experts in Group 1 delivered
higher estimations than those reported in major international cohort studies, estimations from Group 2 were, to varying degrees,
anchored to previous studies. Accordingly, combined distributions for each group and all experts were signicantly dierent, due to
the high sensitivity of the weighted distribution to experts performance in calibration variables. Results of this study suggest that,
while the use of SEJ has great potential for estimating C-R functions for chronic exposure to PM2.5 and premature mortality and its
major sources of uncertainty in countries where no studies are available, its successful implementation is conditioned by a number of
factors, which are analyzed and discussed.
INTRODUCTION
Air pollution is a major risk factor for public health, since public
exposure to air pollution has been linked to numerous negative
health impacts. Over the last two decades, time series and
cohort studies have found evidence of a relationship between
short- and long-term exposure to ne particulate matter
(PM2.5) and increased incidence of adverse health impacts
such as cancer, cardiovascular and respiratory diseases, and
premature mortality.110 In these studies, the relationship
between ambient concentration of air pollutants and the
incidence of a specic health eect is described using
concentrationresponse (C-R) functions. More specically,
results from cohort studies suggest signicant relationships for
long-term PM2.5 exposure and increased mortality from general
and cardiovascular diseases. The most cited studies were
developed in the 90s.11,12 More recent studies have been
performed reanalyzing data from previous work2,1318 and
using new data.1923 While, in general, these studies have
reached some degree of consensus regarding the existence of
this relationship, signicant variability is observed in CR
2014 American Chemical Society
Article
Article
Article
Figure 1. Expert panel estimates for seed variables S1S7. Horizontal lines indicate the true value of each variable, unknown by experts. A complete
description of seed variables is provided in SI.
9721
Based on your interpretation of the evidence, do you want to incorporate a threshold in your characterization of the CR relationship?
What is your estimate of the true percentage change in annual mortality from nonaccidental causes in the adult population of Chile (over 18 years) due to a permanent reduction of 1 g/m3 in annual average
PM2.5 concentrations (over the entire concentration range of 1035 g/m3)? In formulating its response, you must consider eects of reductions in PM2.5 concentrations on both the short and long-term
mortality. To characterize the uncertainty in the CR function, you must provide your estimate as percentiles (fth, 25th, 50th, 75th, and 95th).
Q6
Q7
Note: Question Q4 refers to the shape of the CR function. Questions Q5 and Q6 are related to threshold. Question Q7 is the quantitative question on mortality estimation in Chile.
Do you want to specify if the CR function for PM2.5 and mortality diers across the specied concentration range? If your answer is yes, indicate the form of the functions, the approach to incorporate causality,
and if it considers a concentration threshold.
Do you believe that a concentration threshold for annual nonaccidental mortality associated with population exposure to PM2.5 is detectable in any currently available study
Q5
What evidence do you considered to support or reject the possibility of a causal relationship between reductions in annual average exposure to PM2.5 (including long and short-term exposure reductions) and
changes in mortality in Chile (at annual average 1035 g PM2.5/m3). Do you wish to make a distinction between short-term and long-term relationship? Briey specify your arguments to make such distinction.
Q2
query variables related to causality between PM2.5 exposure and mortality in Chile.
In your opinion, is there a causal relationship between short/long-term exposure to PM2.5 and mortality? The answer must be delivered according to the U.S. EPA standard on weight of evidence for the
determination of causality.24
Q1
Article
RESULTS
Seed Variables. Only Expert B modied his original
estimates after seeing his individual calibration and information
scores. Experts estimates for seed variables are shown in Figure 1.
In general, experts performance in S1, S2, and S3 was good,
although only experts C and D included the real value of these
three seed variables in their estimations. Note that these two
experts provided the largest uncertainty intervals for S1, S2, and S3.
As expected, the entire panel estimated ratios greater than
one for every percentile of seed variables S4, S5, S6, and S7;
nonetheless, overall experts performance in these variables was
low. In general, all experts with the exception of experts C and
E, provided narrow condence intervals for these variables,
missing the real value by large margins.
Query Variables. Causality. All participants agreed that
there is some degree of causality between exposure to PM2.5
and premature mortality: experts B, D, and F considered that
the relationship is causal both in the short and long-term,
experts C and G considered that the relationship is causal for
9722
Article
Figure 2. Uncertainty distributions for the relationship between PM2.5 and nonaccidental mortality in Chile. Box plots represent experts
distributions, the three dierent combination schemes: equal (EW), global (GW) and item weight (IW) for each group (1 and 2) and for all the
experts (A), and the results from four major international cohort studies and the WHO guidelines.
markedly similar, all revealing the same value for the 50th
percentile. The three DM for Group 1 are approximately two
times higher than estimates from Dockery et al.,11 and Laden et
al.,16 while presenting greater uncertainty intervals. The clear
similarity is due to the high weighting of experts C and D in the
combined distribution using the global and item schemes,
estimated based on their performance in the seed variables.
The DM for Group 2, where all experts are equally weighted
(EW-2), received a calibration score of 0.0146. Figure 2 shows
that the combined distributions are closer to results reported by
Pope et al.2,12 On the other hand, the calibration score obtained in
the global weight DM of Group 2 (GW-2) was signicantly lower
than the obtained in the item weight DM of Group 2 (IW-2).
DM distributions combined under the equal, global and item
schemes for the complete panel (EW-A, GW-A, and IW-A,
respectively) present important dierences in their central
estimates (EW-A P50 = 0.9294, GW-A P50 = 2.7830, and IW-A
P50 = 2.4480) as well as in uncertainty intervals. While EW-A
incorporate the results from four major international cohort
studies and the WHO guidelines, the GW-A and IW-A are
similar to the DMs of the Group 1, as experts C and D
obtained the highest calibration scores of the panel.
Robustness Analysis. The robustness analysis for variables
revealed that the calibration score for Group 1 is particularly
sensitive to seed variables S1, S3, and S6, while Group 2, is
more sensitive to the seed variables S4, S5, and S6. For all the
experts, the rst three seed variables (S1, S2, and S3) retained
the greatest inuence on the model.
In relation to the models sensitivity toward experts, the
analysis shows that the calibration score of Group 1 is not
sensitive to the experts and the calibration score of Group 2
showed greater sensibility to Experts B and E.
DISCUSSION
In this study we explore the potential strengths and weaknesses
of SEJ as a way to derive CR functions for regulatory policy
9723
0.001 05
0.000 10
0.000 04
0.000 12
0.000 12
0.01464
0.02825
0.16680
0.011 02
0.004 46
0.39250
0.03583
0.26270
1.340 00
6.610 00
0.783 40
1.994 00
1.882 00
0.582 50
0.833 30
1.105 00
0.583 40
1.671 00
0.325 20
0.523 30
0.553 00
all
variables
1.488 00
7.474 00
0.792 00
2.262 00
2.049 00
0.654 10
0.932 50
1.251 00
0.647 20
1.885 00
0.363 50
0.589 70
0.623 40
seed
variables
relative information
0.001 57
0.000 73
0.000 03
0.000 27
0.000 25
0.009 58
0.007 13
0.008 40
0.142 70
unnormalized
weightc
0.2
0.2
0.2
0.2
0.2
0.5
0.5
0.126 10
0.059 02
0.002 30
0.021 80
0.019 75
0.77110
0.045 07
0.053 08
0.90180
with DM
normalized weight
without
DM
0.026 34
0.001 57
0.000 73
0.000 03
0.000 27
0.000 25
0.021 13
0.007 13
0.008 40
unnormalized
weightc
0.550 70
0.257 80
0.010 05
0.095 21
0.086 26
0.459 20
0.540 80
without
DM
0.90260
0.053 65
0.025 12
0.000 98
0.009 28
0.008 40
0.57630
0.194 50
0.229 10
with DM
normalized weight
global weight
0.208 70
0.001 57
0
0
0.000 27
0.000 25
0.16380
0.007 13
0.008 40
unnormalized
weightc
without
DM
0.99010
0.007 43
0
0
0.001 29
0.001 16
0.91340
0.039 77
0.046 84
with DM
normalized weight
item weight
Three combination schemes are compared: equal, global and item weight. bThe calibration score is the p-value of falsely rejecting the hypothesis that an experts probability statements are statistically
accurate. The maximal value is 1, the minimal value is 0. cRepresents the weight assigned to each expert, calculated as the product of the calibration score (second column) and the information score
relative to seed variables (fourth column). dDM_EW: Equal weight combination scheme. eDM_GW: Global weight combination scheme. fDM_IW: Item weight combination scheme.
Expert A
Expert B
Expert E
Expert F
Expert G
DM_EW2d
DM_GW2e
DM_IW2f
Group 2
Expert C
Expert D
DM_EW1d
DM_GW1e
DM_IW1f
Group 1
expert
calibration
Scoreb
equal weight
Table 2. Results from Using Seed Variables to Obtain the Combined Distribution of the Expert PanelS Estimates for the Mortality Questiona
9724
Article
Article
ASSOCIATED CONTENT
S Supporting Information
*
AUTHOR INFORMATION
Corresponding Author
ACKNOWLEDGMENTS
This research was partially funded by Chiles National Science
and Technology Commission (Conicyt) through the National
Fund for Scientic and Technological Research (Fondecyt,
grant 1130864), and by the National Research Center for
Integrated Natural Disaster Management CONICYT/FONDAP/
15110017.
REFERENCES
Article
(52) Ilabaca, M.; Olaeta, I.; Campos, E.; Villaire, J.; Tellez-Rojo, M.
M.; Romieu, I. Association between levels of fine particulate and
emergency visits for pneumonia and other respiratory illnesses among
children in Santiago, Chile. J. Air Waste Manage. Assoc. 1999, 49 (9),
154163.
(53) Ostro, B. D.; Eskeland, G. S.; Sanchez, J. M.; Feyzioglu, T. Air
pollution and health effects: A study of medical visits among children
in Santiago, Chile. Environ. Health Perspect. 1999, 107 (1), 69.
(54) Roman, A. O.; Prieto, M. J.; Mancilla F, P.; Astudillo O, P.;
Acuna S, C.; Delgado B, I. Aumento del riesgo de consultas