WEBVTT FILE

1
00:00:00.000 --> 00:00:04.350
Ellen and I'm a researcher at the George
Washington University and today I'm

2
00:00:04.350 --> 00:00:07.980
going to be talking about estimating the
health impacts of air pollution using

3
00:00:07.980 --> 00:00:12.599
satellite derived concentrations so
satellite data allows us to estimate

4
00:00:12.600 --> 00:00:19.170
ground level concentrations at
increasingly finer resolutions, ah, these

5
00:00:19.170 --> 00:00:23.160
high-resolution pollution estimates can
then be used to estimate the health

6
00:00:23.160 --> 00:00:27.019
impacts and what neighborhoods within
cities might be at a higher risk of

7
00:00:27.020 --> 00:00:32.520
disease and mortality that's associated
with exposure to air pollution, so in

8
00:00:32.520 --> 00:00:36.210
this study we wanted to estimate the
health impacts of air pollution at the

9
00:00:36.210 --> 00:00:40.860
local level and to do this we look at
Oakland California within the Bay Area

10
00:00:40.860 --> 00:00:45.980
as a case study, so until recently these
type of health impact assessments have

11
00:00:45.989 --> 00:00:53.180
been reported at the global, country or
county level, but this type of reporting

12
00:00:53.190 --> 00:00:58.710
can mask potential spatial heterogeneity
at the local level. So in this study we

13
00:00:58.710 --> 00:01:04.049
varied the pollution sources and rates
of disease in order to see how our

14
00:01:04.049 --> 00:01:10.080
health impacts vary when we use these
varying concentration sources. We hope

15
00:01:10.080 --> 00:01:14.580
that this type of information can inform
actions that cities can take, to take

16
00:01:14.580 --> 00:01:19.670
action to reduce disparities in health
risks attributable to air pollution

17
00:01:19.920 --> 00:01:20.740
Next,

18
00:01:23.520 --> 00:01:26.939
so in order to do this we use
what's called a health impact function

19
00:01:26.939 --> 00:01:32.090
that's derived from epidemiology studies
and these look at linking the health

20
00:01:32.090 --> 00:01:35.240
outcomes of association to air pollution

21
00:01:35.900 --> 00:01:37.860
so the attributable fraction of disease

22
00:01:37.860 --> 00:01:42.060
is the portion of disease that can be
thought of or can be described as

23
00:01:42.060 --> 00:01:48.259
resulting from a particular risk factor
or exposure in this case air pollution

24
00:01:48.259 --> 00:01:52.500
so here I'm going to be talking mostly
about mortality related to air pollution

25
00:01:52.500 --> 00:01:57.540
but we also looked at cardiovascular
mortality as an emergency room visits

26
00:01:57.540 --> 00:02:00.140
and new cases of asthma

27
00:02:06.040 --> 00:02:09.660
sorry guys it's a lot of pressure

28
00:02:12.300 --> 00:02:14.760
okay
so from these epidemiological studies

29
00:02:14.760 --> 00:02:20.580
we're able to derive a risk estimate
that quantitatively assesses the

30
00:02:20.580 --> 00:02:24.630
relationship between a health outcome
and what we would expect for each health

31
00:02:24.630 --> 00:02:27.560
outcome increase in concentration

32
00:02:30.200 --> 00:02:33.080
so in
order to apply this spatially we

33
00:02:33.090 --> 00:02:38.010
actually need a couple more
concentration input or data inputs first

34
00:02:38.010 --> 00:02:42.870
we need the population of our population
of interest and the spatial distribution

35
00:02:42.870 --> 00:02:48.030
of that population we also need the
baseline disease rates or baseline

36
00:02:48.030 --> 00:02:52.410
mortality of that population and the
spatial distribution of that baseline

37
00:02:52.410 --> 00:02:57.150
disease rate and then we also need to
know the pollution concentrations and

38
00:02:57.150 --> 00:03:02.340
the spatial distribution of our paths
pollution concentrations so in this

39
00:03:02.340 --> 00:03:07.320
study we use concentrations from three
sources a fine particulate matter from a

40
00:03:07.320 --> 00:03:14.370
satellite drive model, land..., nitrogen
dioxide from two data sources first a

41
00:03:14.370 --> 00:03:17.960
land-use regression model that
incorporated satellite troposphere at

42
00:03:17.960 --> 00:03:24.020
column estimates of nitrogen dioxides and
nitrogen dioxide from a innovative

43
00:03:24.030 --> 00:03:27.560
mobile monitoring data set where in
Google Street View cars took

44
00:03:27.560 --> 00:03:30.500
measurements of nitrogen dioxide

45
00:03:33.220 --> 00:03:36.000
the
result of this calculation is the number

46
00:03:36.000 --> 00:03:40.200
of cases that we can expect within a
population that could be attributed to

47
00:03:40.200 --> 00:03:41.360
air pollution

48
00:03:45.240 --> 00:03:49.320
so long-term exposure to
fine particulate matter is associated

49
00:03:49.320 --> 00:03:53.790
with increased risk of mortality and it
contributes a large burden of death and

50
00:03:53.790 --> 00:03:58.980
disability in the US and globally so for
our fine particulate matter data set we

51
00:03:58.980 --> 00:04:02.280
used a satellite derived model that was
originally developed at Harvard

52
00:04:02.280 --> 00:04:07.769
University this dataset provided daily
estimates of fine particulate matter at

53
00:04:07.769 --> 00:04:13.760
the one kilometer resolution in the
United States so this was a satellite

54
00:04:13.760 --> 00:04:19.360
derived model, it was a neural network
hybrid model that incorporated aerosol

55
00:04:19.380 --> 00:04:23.819
optical depth from the moderate
resolution imaging spectroradiometer on

56
00:04:23.819 --> 00:04:29.759
the Earth observing satellite it also
incorporated a 3d chemical transport

57
00:04:29.759 --> 00:04:35.219
model central, site monitoring data and
numerous other meteorological and

58
00:04:35.219 --> 00:04:41.520
land-use variables the result is that we
can see the estimates show that within

59
00:04:41.520 --> 00:04:44.999
the Bay Area the daily maximum
concentrations for fine particulate

60
00:04:44.999 --> 00:04:50.129
matter regularly exceeded the US
Environmental Protection Agency's daily

61
00:04:50.129 --> 00:04:54.500
threshold of 35 micrograms per meter
cubed

62
00:04:55.460 --> 00:05:00.360
so our second concentration of interest
is nitrogen dioxide and nitrogen dioxide

63
00:05:00.360 --> 00:05:04.860
is a traffic-related air pollutant
exposures which has been associated with

64
00:05:04.860 --> 00:05:09.979
increased asthma incidents and increased
mortality so we wanted to compare how to

65
00:05:09.979 --> 00:05:15.319
varying NO2 concentration datasets
influenced our health impact estimates

66
00:05:15.319 --> 00:05:22.439
so for our first model for NO2 we used
a land-use regression model and land-use

67
00:05:22.439 --> 00:05:27.060
regression basically just takes land-use
characteristics and other model inputs

68
00:05:27.060 --> 00:05:32.279
to estimate data where we don't have
measured concentrations so this

69
00:05:32.279 --> 00:05:35.939
particular model developed to Oregon
State University incorporated satellite

70
00:05:35.939 --> 00:05:40.860
troposphere at column estimates of
nitrogen dioxides from the DAICHI

71
00:05:40.860 --> 00:05:47.520
instrument and the Gomi 2 instrument
the researchers found that the satellite

72
00:05:47.520 --> 00:05:51.180
column concentrations and measures of
road density were the largest

73
00:05:51.180 --> 00:05:55.040
contributors to NO2 ground level
concentrations

74
00:05:58.460 --> 00:05:59.939
so in order to compare

75
00:05:59.939 --> 00:06:04.649
our land-use regression estimates we
wanted to use measure data of NO2 and

76
00:06:04.649 --> 00:06:09.689
for this we used a innovative mobile
monitoring technique wherein Google

77
00:06:09.689 --> 00:06:14.159
Street View cars took repeated 30 meter
Road segment measurements of NO2 in

78
00:06:14.159 --> 00:06:18.360
neighborhoods throughout the Bay Area
and for Oakland this is Oakland

79
00:06:18.360 --> 00:06:23.089
California this resulted in three
million, over three million data points

80
00:06:23.089 --> 00:06:28.319
which was then aggregated to an annual
average and then we aggregated to 100

81
00:06:28.319 --> 00:06:30.830
meters here

82
00:06:31.960 --> 00:06:36.500
the Google Street  data sets
concentrations found peak concentrations

83
00:06:36.500 --> 00:06:41.810
near roadways and what we can also see
is that when we compare the land-use

84
00:06:41.810 --> 00:06:45.560
regression to the Google Streetview the
land-use regression had a higher median

85
00:06:45.560 --> 00:06:49.819
concentration of values but the Google
Street View concentrations had a larger

86
00:06:49.819 --> 00:06:55.610
range of values owing to those peak
concentrations near roadways so we

87
00:06:55.610 --> 00:07:00.259
hypothesized that this highly resolved
spatial data near roadways in

88
00:07:00.259 --> 00:07:04.520
combination with highly resolved
baseline disease rates would allow us to

89
00:07:04.520 --> 00:07:09.880
identify disparities in health burden
attributable to air pollution within cities

90
00:07:15.520 --> 00:07:19.660
so the other portion of our
project was to look at how baseline

91
00:07:19.669 --> 00:07:23.840
disease rates influenced the risks
attributable to air pollution within

92
00:07:23.840 --> 00:07:30.439
cities so County baseline disease rates
are pretty easy to obtain but they can

93
00:07:30.439 --> 00:07:34.909
mask potential spatial heterogeneity for
neighborhoods so for example if you

94
00:07:34.909 --> 00:07:39.110
apply a County based on disease rate to
entire neighborhood you might or the

95
00:07:39.110 --> 00:07:42.889
entire county you might overestimate the
baseline disease in some areas and

96
00:07:42.889 --> 00:07:47.599
underestimate it in others so for our
study we obtained census block group or

97
00:07:47.599 --> 00:07:51.529
neighborhood level based on disease
rates and what we found is when we

98
00:07:51.529 --> 00:07:54.770
compared county and neighborhood level
based on disease rates in our health

99
00:07:54.770 --> 00:08:00.919
impact function was the census block
group yielded a higher median burden of

100
00:08:00.919 --> 00:08:05.659
disease as seen by the box plots and it
also yield the different spatial

101
00:08:05.659 --> 00:08:11.210
patterns as compared to the county rates
we can see in the maps, so from this we

102
00:08:11.210 --> 00:08:14.930
concluded that without this type of
highly resolved neighborhood baseline

103
00:08:14.930 --> 00:08:20.870
disease rates we might not adequately
identify disparities in the health

104
00:08:20.870 --> 00:08:24.560
burden attributable to air pollution
within cities

105
00:08:30.580 --> 00:08:35.080
so the other portion of our study was of
course to compare how the satellite

106
00:08:35.089 --> 00:08:38.510
drive Land-use regression model compared
to our Google Streetview

107
00:08:38.510 --> 00:08:42.680
data set and what we found is that the
land-use regression model yielded higher

108
00:08:42.680 --> 00:08:47.450
estimates of mortality attributable to
air pollution and that was owing to the

109
00:08:47.450 --> 00:08:51.950
higher concentrations of the land-use regression model we did also find that the

110
00:08:51.950 --> 00:08:55.370
land-use regression and the Google
Streetview yielded fairly similar

111
00:08:55.370 --> 00:08:59.750
overall spatial patterns but the Google
Street View data set was able to

112
00:08:59.750 --> 00:09:05.660
identify a larger magnitude of spatial
heterogeneity within cities particularly

113
00:09:05.660 --> 00:09:11.660
with hotspots near high-density roadways
such as this is I 880 in Oakland California

114
00:09:16.860 --> 00:09:21.440
so in this presentation we've
seen how highly resolved spatial data

115
00:09:21.440 --> 00:09:26.540
can help Public Health's researchers
identify areas within cities that are

116
00:09:26.540 --> 00:09:32.900
experiencing higher rates of disease
attributable to air pollution the data

117
00:09:32.900 --> 00:09:35.770
sources we explored were satellite
derived concentration estimates

118
00:09:35.770 --> 00:09:40.720
innovative mobile monitoring and highly
resolved based on disease rates

119
00:09:40.720 --> 00:09:45.589
particularly we demonstrated that the
county and that the more highly resolved

120
00:09:45.589 --> 00:09:50.959
baseline disease rates did a good job of
letting us see where these disparities

121
00:09:50.959 --> 00:09:54.680
within cities exist so if we don't have
that highly resolved baseline to deta

122
00:09:54.680 --> 00:09:57.680
data we might not adequately identify
these

123
00:09:59.180 --> 00:10:00.640
so just to kind of bring this all

124
00:10:00.650 --> 00:10:05.390
together at a spring 2019 meeting at the
George Washington University we got

125
00:10:05.390 --> 00:10:09.800
together local and state representatives
of different governments and they really

126
00:10:09.800 --> 00:10:14.450
emphasized the need for this type of
highly localized data in order for them

127
00:10:14.450 --> 00:10:18.950
to inform their decision-making on the
local level so it's our hope that this

128
00:10:18.950 --> 00:10:24.140
type of local health impact assessment
can be used in cities worldwide to help

129
00:10:24.140 --> 00:10:26.540
inform local decision-making

130
00:10:29.720 --> 00:10:32.920
so that I'd
just like to thank our in collaborator

131
00:10:32.930 --> 00:10:36.920
collaborators at the Environmental
Defense Fund, University of Texas Austin

132
00:10:36.920 --> 00:10:41.390
Harvard University, the Alameda County
Public Health Department and Oregon

133
00:10:41.390 --> 00:10:44.600
State University and of course
generous funding from the Environmental

134
00:10:44.600 --> 00:10:47.630
Defense Fund and NASA for funding our
group at the George Washington University