Climate Change and Health Seminar: "Climate, Energy, & Inequity: From Exposures to Epidemiology"

July 29, 2021

Information

June 30, 2021
Dr. Daniel Carrión, Postdoctoral fellow at the Icahn School of Medicine at Mount Sinai

Dr. Daniel Carrión joined YCCCH for this special seminar on household energy, energy insecurity, air pollution and why/how it connects to climate & health.

ID6840

To CiteDCA Citation Guide

00:07<v ->Welcome to this special seminar</v>
00:11being sponsored by the Yale Center
00:13on Climate Change in Health.
00:15And it's a pleasure to welcome Daniel Carrión
00:23who is currently a postdoctoral Fellow
00:26in Environmental Medicine and Public Health
00:28at the Icahn School of Medicine at Mount Sinai.
00:33He received his PhD from
00:35Columbia Mailman School of Public Health
00:39from the Department of Environmental Health Sciences
00:43and it was in their Climate and Health Program,
00:45which is a really great program that our own Chi Chen
00:52has been closely associated with in the past.
00:56And so we're really looking forward to Daniel's presentation
01:00on Climate, Energy and Inequity:
01:03from Exposures to Epidemiology.
01:05So, Daniel, welcome.
01:08<v ->Thank you so much.</v>
01:10So I'm excited to speak to you all today
01:13and just by way of a little bit more introduction,
01:17I completed my BA at Ithaca College in 2008,
01:22and if you remember 2008,
01:25that was right when the global recession happened,
01:28so a great time to graduate from college.
01:31So I had two part-time jobs, one where I was working
01:34actually for the Health Department in Tompkins County
01:36in New York state, and the other one where I was working
01:40for the Solid Waste Division where I was doing composting
01:42and recycling education and outreach.
01:45I then ended up leaving
01:47and going to Hudson River Healthcare,
01:48which is a network of federally qualified health centers
01:53across New York state, about 25 at the time,
01:57where I was helping manage outreach
02:02and programming for folks with HIV, folks who were homeless,
02:06folks in public housing, and migrant farm workers.
02:09I was concurrently doing my masters in public health
02:12and environmental health sciences
02:13at New York Medical College.
02:15And then after completing my MPH
02:18ended up leaving to go to Columbia university
02:21where I started a pipeline program called
02:24the Summer Public Health Scholars program,
02:26a CDC funded program to increase the diversity
02:30of the public health workforce,
02:32specifically around health equity.
02:35I then started my PhD as Rob mentioned
02:38in the department of environmental health sciences
02:42and the climate and health program
02:44and completed that in 2019.
02:47And then now at the Icahn School of Medicine as a post-doc.
02:51And so I'm excited to tell you about the work that I've done
02:56in the most recent part of this journey,
02:59which I characterize being at this nexus of climate energy
03:03and health inequity.
03:05So we all know that energy lies at the source of our climate
03:09crisis, societal decisions on where we derive energy,
03:13how much we need and what we use it for
03:19are all leading to increasing global temperatures
03:23that we have been observing and we'll continue to see.
03:28But we've run a dynamic tension here, right?
03:31Because energy is fundamental to public health.
03:35It's fundamental for folks to stay healthy,
03:38from the energy that we use to cook with,
03:41to the energy that we use in the winter to stay warm,
03:44to the energy that we use in the summer to stay cool,
03:49we need energy.
03:51And so I've been fortunate to work in all three
03:53of these spaces, thinking about this,
03:56these energy tensions in public health,
03:59but for the scope of this talk,
04:00I'm going to only tell you about two of them,
04:03which is about my work in household energy and air
04:05pollution related to cooking,
04:07and then more recently temperature epidemiology
04:10from summertime temperatures.
04:14So quickly about my dissertation work and household energy
04:18and air pollution in low and middle income countries.
04:22As background, 3 billion people around the world
04:28experience energy poverty,
04:30which is characterized by cooking and or heating with wood,
04:34dung, charcoal, or other biomass fuels.
04:40And although the proportion is decreasing overall
04:45of the population that relies on these fuels
04:49because of population growth,
04:51the absolute counts are actually increasing and the highest
04:55increases are actually in Sub-Saharan Africa.
04:59And so this is the stove that you would see in many parts
05:03of Sub-Saharan Africa, it's called the three stone fire,
05:08which you might guess because there are three stones
05:11that prop up a pot and underneath biomass is combusted.
05:18We're concerned about this because the combustion of that
05:21biomass leads to a mixture of compounds collectively
05:26referred to as household air pollution.
05:29And so that comprises CO2 particulate matter,
05:34carbon monoxide,
05:35polycyclic aromatic hydrocarbons amongst others,
05:40and both the deforestation associated with biomass
05:43harvesting depending on country and the combustion
05:47are projected to actually contribute to climate change.
05:52And we also know that exposure to household air pollution
05:56is associated premature deaths each year,
06:00millions of premature deaths each year,
06:03the largest proportion from lower respiratory infections.
06:07And you might know that lower respiratory infections
06:10are actually the leading killer of children
06:13under five in lower and middle income countries.
06:16And so it's widely agreed that the solution here
06:19is to scale up cleaner cooking alternatives
06:23like liquified petroleum, gas, electric, and induction.
06:30And in Ghana, as in many other countries,
06:33LPG represents the cheapest and most accessible options
06:37of the three that I just mentioned because the other two
06:42electric and induction requires stable and extensive
06:46electricity grids that don't exist
06:48in many parts of the world.
06:50But if you're unfamiliar with this literature,
06:52I would understand if some folks in the audience
06:55are confused at how using a fossil fuel can actually help us
06:59fight climate change.
07:04The atmospheric science behind this is complicated
07:06and outside the scope of my talk today,
07:09but rest assured that the international panel on climate
07:12change indicates that activities consistent
07:16with the greenhouse gas emission reductions needed
07:19for a warming of 1.5 degrees Celsius
07:24world includes transitions to clean cookstoves
07:29that are gas based or electric based.
07:33And unfortunately, atmosphere projections
07:37that are Ghana-specific are actually unavailable
07:40at the moment, but one done in Cameroon
07:45undergoing a similar LPG transition
07:49shows that there are projected net cooling benefits
07:54of switching to LPG rather than continued use
07:58of biomass fuels.
08:00And so this then represents in many parts of the world
08:04climate mitigation opportunity
08:07with potential health co-benefits.
08:12And so my thesis works set out to try and provide evidence
08:15to support clean cooking efforts.
08:18The relationship between energy,
08:20poverty and disease can be described as a pathway
08:23from poverty to energy poverty,
08:27which then causes household air pollution,
08:30and then the exposure to that household air pollution
08:33leads to a whole host of diseases.
08:36And there are particularly three parts of this pathway
08:39that we can try to interrupt in this relationship
08:44between poverty and disease in this context.
08:47So we can focus on making the clean available,
08:51which is a moniker from the late Kirk Smith,
08:55essentially saying identifying interventions
08:58to increase the uptake of clean cookstoves
09:00like induction or LPG.
09:04We could interrupt this part of the pathway,
09:06which is to make the available clean by identifying ways
09:11to reduce exposures from biomass combustion,
09:15such as improved cookstoves that have interventions
09:19like increasing ventilation,
09:21thereby increasing the efficiency of combustion.
09:26And then finally, we can do health research
09:30to understand biological pathways
09:32for improved treatments or interventions.
09:38My work was particularly focused on these two parts
09:41of the pathway, and I'll quickly sum up my dissertation
09:46in one slide, which is the first paper
09:51in my dissertation was where I
09:54created a new framework to try and understand why recipients
09:58of new cookstoves often end up
10:01stopping using those cookstoves
10:04and we refer to this as stove use discontinuance.
10:09Acknowledging that a lot of people who receive
10:11new cookstoves end up stopping their use in the longer term,
10:17we ended up then trying to design an intervention
10:21to support a government effort.
10:24So the government actually freely distributes LPG stoves
10:29in rural areas in Ghana.
10:31And so we designed and implemented an intervention
10:36to try and increase the long-term use of those stoves.
10:41The findings suggest that more fundamental policy changes
10:44are actually needed just rather than a simple intervention.
10:49And finally understanding biological pathways
10:54from data from a cohort study,
10:56we used banked nasal swabs from infants
11:00of the age of one or less
11:03and found that household air pollution is associated
11:06with increased presence of bacterial and not viral microbes.
11:11And this is important because there's other literature
11:14that otherwise indicates that household air pollution may be
11:17contributing to bacterial forms of pneumonia
11:21and not viral forms of pneumonia
11:23and so this is trying to understand that ideological pathway
11:27a little bit better.
11:31So with that very brief overview of my thesis work,
11:35I wanted to spend more time on my current portfolio,
11:39which is focused on ambient temperature,
11:41temperature epidemiology, and energy insecurity.
11:47And the motivation here is simple.
11:51We're living it right now.
11:52Climate change means that there's an increased frequency
11:56and intensity of extreme heat events
11:59and hotter average summers.
12:01And we know that those higher temperatures are associated
12:05with a whole host of health outcomes from cardiovascular
12:08to respiratory, to renal,
12:11to even violence and other non-health outcomes,
12:15but still very health relevant like educational performance.
12:20And there's also work that shows that increased ambient
12:24temperatures are associated with perinatal outcomes
12:28like pre-term birth.
12:30And there's an important opportunity here
12:32because temperature epi has been largely focused
12:35on older adult populations
12:37and so there's an opportunity to grow the literature
12:41thinking about pediatric populations.
12:47So I first want to tell you about a study
12:49that we're wrapping up right now,
12:51thinking about the case process over design as a way
12:55of studying the relationship between ambient temperature
12:58and preterm birth.
13:04And the motivation here
13:06I think is also simple for a public health crowd.
13:10That preterm birth is a major health outcome
13:16that's associated with high infant mortality.
13:19It's also one of the most pronounced and persistent
13:23racial disparities that we know of,
13:26and it not only represents poor health
13:29potentially in the immediacy of birth,
13:33but also potentially a trajectory of poor health
13:37in the longterm.
13:39Many of the health outcomes are also health disparities
13:43for communities of color.
13:45And there's a growing literature on the relationship
13:48between ambient temperature and preterm birth.
13:51One of the initial studies identifying this association
13:54was actually from Bosu at all in 2010,
13:58a study based in California using
14:00the case crossover study design.
14:04So if you're unfamiliar with the case crossover study
14:07design, a quick introduction.
14:09It's a case-only study design that compares the case time
14:15to control times when the event did not happen.
14:20And it's been widely used in air pollution epidemiology
14:23and is increasingly used in temperature epidemiology.
14:28It's a temporal comparison,
14:32meaning that it's comparing the same person to themselves
14:35at different time points.
14:37And so a real perk there is that then it's not vulnerable
14:41to person level forms of confounding.
14:47However, proper control selection is then pivotal for proper
14:53inference because you want to make sure that you
14:55are controlling for potential temporal confounders
14:59and other temporal forms of bias.
15:02And a key assumption of this design is that there are no
15:07trends in the risk of the outcome over time.
15:11And it was actually pointed out in a commentary
15:14from that original Bosu paper that I mentioned
15:18that preterm birth actually violates this assumption.
15:22And this should be pretty intuitive to folks in the audience
15:25because the risk of birth changes
15:30pretty secularly over gestation.
15:33And so this is something that we need to think about
15:35if we're using this study design
15:38for ambient environmental exposures.
15:45However six other studies have employed this study design
15:49for preterm birth since 2010,
15:52specifically for ambient temperature that we're aware of.
15:56And I'm sure that number is much higher
15:58if we also consider air pollution.
16:06So that this was a great opportunity for a simulation study.
16:10So for those who are unfamiliar,
16:12a simulation study are essentially computational experiments
16:15where we can test the behavior of our epidemiological
16:19studies under controlled circumstances.
16:23So first what we do is we create a dataset and then we embed
16:28a known association in that dataset.
16:32We then test our epidemiological analysis'
16:35ability to recover that association.
16:38Then we try to repeat,
16:40or we repeat this a thousand times to represent
16:43some of the stochasticity of the underlying distribution.
16:47And then we could see if different strategies
16:49or specifications of models
16:51can actually improve our inference.
16:56More specific, what data did I use to do this?
16:59Well, LaGuardia Airport has temperature data
17:03readily available for download online.
17:06So we downloaded LaGuardia temperature data
17:09as our exposure data.
17:12And then for our health data, we actually downloaded CDC
17:16wonder data to create estimates of daily preterm births
17:22by gestational age from 20 to 36 weeks.
17:28And just as a quick definitional thing,
17:30preterm birth is generally a birth that take place
17:34before 37 weeks.
17:37We got these data for 2007 and 2018 from,
17:42and then we created data sets with a range
17:46of simulated effects ranging from 0.9 to 1.25.
17:51I don't think anyone thinks that temperature
17:53is protective of preterm birth,
17:56but we wanted to see how malleable these models
18:01were to different underlying assumptions.
18:06And then we do these case crossovers to see how our model
18:10does at recovering the simulated effects.
18:14We ended up doing this using a time stratified control
18:18selection for three different time periods.
18:21So we did it for a two week time stratified,
18:24a 28 day time stratified, and a month time stratified.
18:29And we limit our case crossover to warm month analyses,
18:33which is consistent with other studies in this literature.
18:38And again, we do this a thousand times to kind of represent
18:41some of that stochasticity of the underlying distrobution.
18:46So these are the input data that we use.
18:50So up here are, is the temperature data
18:55from LaGuardia Airport
18:58and down here are the estimated number of births
19:01on a given day that we used from the CDC wonder database.
19:06And then this orange region
19:08is the warm month time period that we used.
19:15So the main result that I'm showing you here
19:17is for absolute bias.
19:20And so absolute bias is simply the difference between
19:23the simulated relative risk with the coefficient that we get
19:27from the case crossover in the log scale.
19:31And I'm showing you first a relative risk of one,
19:34meaning that there's no association between temperature
19:38and preterm birth.
19:40And you could see that using all three of these study
19:43designs, we actually get relatively unbiased results
19:48with the medians hovering around zero.
19:54If we look across the entire range of our embedded effects,
19:58we see relatively consistent results where all three
20:02of these case control selection designs actually yield
20:07relatively unbiased results, with our two-week stratified,
20:12yielding the noisiest results characterized here
20:16by a wider intercore tile range.
20:21And then when we looked at coverage,
20:23so coverage would be the coverage
20:26of the 95% confidence intervals.
20:29What percentage of the time does the confidence interval
20:33actually include the true embedded effect?
20:36And you would hope for a model that that would be
20:40consistently 95% of the time.
20:41And indeed we see that these models are relatively stable
20:45with approximately 95% at all of these risks embedded.
20:55So this is really important work because this shores up
21:00the evidence that we have
21:02for the case crossover study design
21:05and ambient exposures and preterm birth,
21:10which I think is really important.
21:12We ended up doing 24,000 simulations and corresponding
21:15case crossovers, finding that the models
21:18are relatively unbiased.
21:20And we're excited about wrapping up this project
21:24because we've tried to enhance reproducibility
21:27of our findings and results by using the targets package
21:31in R, which then means that other folks
21:37can go and rerun these analyses and can actually swap out
21:42different years or regions and their analysis,
21:45which aids an extensibility of this analysis.
21:50And now we're actually using the case crossover analysis
21:56to think about a national level analysis
21:59that we're doing actually in Mexico
22:02and hopefully future studies in the U.S. as well.
22:08But much the same way that we're thinking about
22:11epidemiological methods, we're also thinking about improving
22:15our exposure methods.
22:17And so here, I want to tell you about a project
22:19that we just published on,
22:21thinking about a one kilometer hourly air temperature model
22:24across the Northeastern United States
22:27from Maine to Virginia
22:29and this is fusing ground data
22:33with satellite remote sensing data.
22:37And the inspiration for me here is that there is a small,
22:43but growing literature on temperature disparities,
22:47that temperature is perhaps unevenly experienced
22:52based on race, ethnicity, income,
22:56and other forms of potential vulnerability.
23:00And so one limitation, however,
23:04with some of these past studies is that they either use land
23:08surface temperature, which is remotely sensed
23:10with satellites and related to air temperature,
23:14but not exactly air temperature,
23:16or they use forms of land cover,
23:21and land use that are associated with temperature,
23:24but again, not empirical measures of temperature
23:28and so an opportunity then to try and grow this literature,
23:32thinking about these potential temperature disparities.
23:39So the goal here is to create this one kilometer
23:42hourly air temperature model
23:44to be able to produce predictions
23:46between the time period of 2003 to 2019.
23:51So we ended up using national oceanic,
23:55atmospheric and atmospheric administration data
24:00for ground stations throughout this region
24:03as our ground truths for air temperature.
24:05And so that's what's depicted in red
24:08across our study region.
24:10These are the locations of all of the ground sensors
24:13that we used in our model.
24:16We then collected 34 predictors that we thought
24:21would help us characterize the spatial and temporal patterns
24:24of cooling and heating throughout the day.
24:29And the goal here is to be able to create consistent
24:32and reliable predictions of air temperature across
24:36this region, even in places
24:38that we don't have ground observations.
24:46So we tested five different statistical approaches
24:50to actually create these predictions
24:55and show their differences in performance in our paper.
24:59For the sake of time,
25:01I'm just going to tell you the punchline,
25:03which is that we ended up using the XG boost model
25:07for our final predictions.
25:09So the XG boost model is a powerful machine learning model
25:14that we used and had to adapt to create
25:19a spatial temporal predictions.
25:23And what we ended up doing was actually comparing
25:27our XG boost model to the NLDAS-2 model.
25:32So NLDAS-2, if you're unfamiliar is a NASA product
25:37that also gives hourly predictions
25:39and it's what the CDC uses for their heat and health
25:43tracking system, as well as some of their research.
25:48And so we thought that this was an important model
25:51to benchmark again.
25:56So these are the predictions from our XG boost model,
25:59from the hottest midnight of our data set, July 22nd, 2011.
26:05And so you can see across this Northeast region
26:09from Virginia to Maine,
26:11that we reconstruct a great deal of spatial heterogeneity.
26:17Again, this is for one hour,
26:19the highest midnight of our time period.
26:23And when we zoom in to a sub region,
26:29this, in this case being New York City,
26:34we see that we reconstruct a great deal of spatial
26:37heterogeneity from the urban heat island effect.
26:42And I should have mentioned earlier,
26:45I mentioned that NLDAS-2 is hourly,
26:49but it's actually at a much coarser spatial resolution.
26:53So these larger grid cells overlaid our predictions
26:58are actually the NLDAS-2 grid cells.
27:02And it's important to note here that in this one,
27:06NLDAS-2 grid cell, you have most of Manhattan,
27:10a big chunk of the Bronx and a little bit of Queens
27:14that would get one prediction for all of that region,
27:18with the NLDAS-2 predictions,
27:21but we can reconstruct a great deal of heterogeneity
27:25within that region.
27:29And we think that that then is related
27:32to the performance of these models.
27:35So these are the root mean squared errors from just 2019
27:40from our XG boost model versus the NLDAS-2 model.
27:46So RMSE is a measure of predictive accuracy and the goal
27:51is to have lower RMSEs.
27:54And so we show that our model
27:56has a low RMSE of 1.4 Celsius,
28:01whereas the NLDAS-2 model has a RMSE of 2.4 Celsius.
28:09When we look across the entire region across all years,
28:14we see that the XG boost predictions have one third
28:17of the mean squared error of the NLDAS-2 predictions.
28:26But given the small literature on temperature disparities,
28:29we were curious to see if our model was also associated
28:33with a measure of social vulnerability.
28:37And so what we decided to do was actually conduct a limited
28:42application to look at the relationship between our model
28:46and the NLDAS-2 model with social vulnerability.
28:52So what we did was we used the CDCs social vulnerability
28:56index, which are a composite of 15 census variables,
29:02including socioeconomic status, housing, transportation,
29:07language isolation, amongst other characteristics.
29:12And these are variables that the CDC uses to identify
29:16communities that may need support before,
29:19during or after a disaster.
29:23The results from the social vulnerability index
29:26are proportional.
29:28It produces measures from zero to one.
29:31And so we decided to use mixed models
29:37to associate our XG boost model and the NLDAS model
29:43with social vulnerability to see how they were associated
29:47with social vulnerability at the census tract level.
29:51We wanted this to be a limited application
29:54so we only did it for one hour of one day from that hottest
29:58midnight that I showed you earlier.
30:04And here are the results.
30:07So, as I mentioned earlier,
30:09the CDC social vulnerability index is a proportional measure
30:12from zero to one.
30:14And so for a unit increase of the CDC SVI,
30:20we see that the NLDAS-2 model shows an increase
30:24of temperature of 0.18 Celsius.
30:28However, when we look at the XG boost model,
30:30we see that our model has a stronger relationship
30:35with an increase in temperature,
30:37average temperature of 0.71 Celsius.
30:44And just to ground that in some places that you might know,
30:49so if we look at New York City,
30:52two boroughs of New York City, Manhattan and the Bronx,
30:57and then we look at two counties in upstate New York,
31:02you would see that the NLDAS-2 model has a very,
31:07very shallow gradient of temperature and social
31:11vulnerability across these temperature predictions.
31:16However, with our XG boost model,
31:18because we reconstruct much more spatial heterogeneity,
31:23we see much more of a strong relationship
31:28with the social vulnerability index.
31:32So with the caveat that this is one hour of one day,
31:36what this implies to us is that there's potentially exposure
31:40misclassification in coarser models.
31:44And that that exposure misclassification may be differential
31:48by neighborhood vulnerability.
31:53So as a takeaway here,
31:56we've created highly accurate air temperature predictions
32:00that we think are right for application
32:03to social science, exposure science,
32:07and epidemiological studies.
32:11But wait, there's more,
32:13I think that this is a great segue
32:16because I'm currently expanding on these questions
32:19with work that I'm doing at the moment.
32:23And so right now, I want to quickly tell you about work
32:27that I have underway to try and explore these exposure
32:31disparities further and point to its potential importance
32:36for epidemiological methods.
32:39And so this is about thinking about residential segregation,
32:43air temperature, and circulatory mortality.
32:49So for the first part of the analysis,
32:51I'll be looking at exposure disparities,
32:54similar to the methods that I just showed you,
32:57but with some key differences.
33:00So unlike the last analysis,
33:03this time I actually want to look at the differences
33:07and the predictions by race.
33:09We know that we have suggestions from past literature
33:13that there are differences in exposure by race
33:16and ethnicity and so we want to look at this
33:20by race and ethnicity as well
33:25now that we have air temperature predictions.
33:28And so what we decided it had to do was we decided to
33:32aggregate our models to the census tract level
33:35like we did before and then we wanted to see what
33:40the differences were potentially in an experienced summer.
33:46And so what I did was I wanted to compare are the summertime
33:50aggregates so I borrowed from the energy literature
33:54and computed cooling degree days.
33:58So if you're unfamiliar with cooling degree days,
34:01generally speaking, what it is is measures
34:04of how much hotter a day is than a threshold value.
34:08Generally in the U.S., the threshold value that's used
34:12is 65 degrees Fahrenheit, or 18.3 degrees Celsius.
34:18So, as an example, if today is 67,
34:22which I wish that it were, but if it were 67 outside today,
34:26that would give us two cooling degree days.
34:29And then you repeat that for every other day,
34:31and then add up all of those cooling degree days
34:35for the summertime values.
34:39For now I'm only conducting a comparison
34:42of exposure experiences by black and white people,
34:46but in the future, I want to consider more racial groups
34:51to try and characterize these exposure disparities better.
34:57And you can imagine that if we see differences by race,
35:02someone could make an argument that it might be
35:06because different people live
35:09in different parts of the region.
35:11So for example, saying that more white folks live
35:15in the Northern most parts of the region like Maine
35:20and more black folks live in the Southern most part
35:22of the region like Virginia.
35:25And so we wanted to then make this within county comparison
35:31within geographic compact geographies,
35:35to look at exposure disparities within these
35:40more relevant administrative units.
35:43And so to address that,
35:45we then took a similar approach of comparing tracks
35:49within counties with our predictor variable,
35:53being the proportion of the census tract
35:58that was comprised of black folks,
36:00and then using random intercepts and slopes by county
36:05to then get county level comparisons.
36:11On the epidemiological side of things,
36:13you can imagine that getting health data that covers
36:16the entirety of this region is pretty difficult
36:20so we use it as an opportunity to get creative.
36:23We, again, access to CDC wonder data
36:27and although I'm interested in child health,
36:30CDC wonder data has some major limitations
36:34if we're thinking about a rarer health outcome
36:37like preterm birth.
36:40Data are provided are at very coarse geographies.
36:44In this case, data are only provided at the county level,
36:48and they're also only provided for course time spans.
36:52And then data that are counts that are below 10
36:58are suppressed for privacy concerns.
37:02So, because CVD mortality is a much more common event,
37:07we decided to conduct this analysis with CVD mortality.
37:12There are still however,
37:14a fair amount of suppressions of data
37:18and so to deal with that,
37:19we ended up using a left censored Poisson regression
37:24since there would be left censoring for lower counts.
37:29And really one of the things that I'm getting at here is
37:32around this question of exposure misclassification.
37:37So for example, in many environmental epidemiology studies,
37:40there's oftentimes an analysis that looks at effect
37:43modification by race, often finding higher effect estimates
37:47based on race and ethnicity.
37:49And while there are sometimes reasons to think that this
37:52might be the case, depending on exposure and context,
37:56I am often left wondering if it's potentially a consequence
38:01of underlying exposure disparities that our exposure models
38:05are not picking up.
38:08And so with that inspiration,
38:10I ended up doing four different regressions,
38:15two regressions for white folks using both exposure models
38:20and two regressions for black folks using both regression
38:23models or prediction models, I should say.
38:27And since this ended up being at the county level,
38:29I tried to preserve some of the exposure differences
38:33by computing weighted by track level racial composition,
38:39aggregated up to the county level.
38:47So these are preliminary results just for the year 2019.
38:53So this plot is simply looking at the distributions by race
38:57across the 13 states including DC.
39:02And what we see here is that actually both models
39:05appear to reconstruct a temperature disparity
39:09between whites and blacks.
39:12However, our XG boost model has a much more smoothed out
39:18distribution for black folks.
39:24And when we actually look at the median values experienced,
39:28we see that they're about the same for white folks,
39:32but between these two prediction models.
39:34But in fact, we have higher exposures for black folks
39:38with our XG boost model.
39:41But this is just looking across the entire region,
39:44this isn't actually of the results from our analysis
39:47and so from that linear mixed effect model
39:50that I mentioned earlier,
39:52we look to see at how these were related to the proportion
39:56of black people living inside of a census tract
40:00and we found that a zero to one increase for the proportion
40:06of folks was associated with 25 higher cooling degree days
40:11for the NLDS to model.
40:14But for the XG boost model,
40:16we reconstruct approximately 68 cooling degree days.
40:23And so we think that this is potentially important
40:27for reconstructing some of these potential
40:30exposure disparities and on the epidemiological
40:34side of things, when we do a stratified model
40:38for white folks, we see a modest but significant effect
40:42of approximately 1.04.
40:45But when we look at those as effect estimates
40:48for black folks, we see much higher effect estimates
40:53for both models.
40:54However, this is for the NLDAS-2 model with about 1.24
41:00as the effect estimate.
41:04It was mentioned in the slide
41:05but I should've said it before,
41:07these are scaled per 92 cooling degree days
41:11or one cooling degree day average increase
41:15across our time span.
41:18And so for the XG boost model,
41:21we see that we get a much lower,
41:25but still higher than for whites effect estimate of 1.14.
41:32So what this means to me,
41:34or implies to me that there is potentially exposure
41:39misclassification that can appear
41:44in epi models as greater susceptibility.
41:49And so I think that there is an opportunity here to think
41:52further about these models and what they can lend us
41:56for health disparities types of research.
42:02So some next steps here is that I have data for more years
42:07than just 2019, so I'm going to include more years
42:10in this analysis.
42:12We also know that there are exposure disparities
42:16for other forms of environmental contaminants
42:20like ozone or PM2.5.
42:23And so I want to potentially control for these
42:25as spatial temporal confounders,
42:29potentially contributing to these relationships.
42:33And then I want to include explicit measures of segregation.
42:38So, as I mentioned, I showed the proportion of black folks,
42:43but there's a whole host of literature that actually shows
42:46different measures of segregation like the dissimilarities
42:50index or the index of concentration at the extremes.
42:54And I would like to use these
42:56as potential predictors in these models.
43:02And then finally, I want to analyze these disparities
43:06in relation to energy data
43:09because I'm interested in studying some quantitative
43:15research between energy burden and energy insecurity,
43:20which leads me to some of my
43:21future directions and opportunities.
43:26So if you're unfamiliar with energy insecurity,
43:29this is a relatively new framework that my colleague
43:33Diana Hernandez at Columbia has used and described
43:40as a framework that outlines the interplay
43:42between energy needs, financial constraints,
43:46and behavioral adaptations.
43:48So I think a lot of us are familiar with this concept
43:54in what's referred to as the heat or eat dilemma.
43:58So the heat or eat dilemma describes the kind of precarious
44:01situation that historically poor families have been put in
44:06of during the winter time,
44:09do they keep themselves warm or do they forgo some staples,
44:14like a healthy meal, or perhaps they get their heating
44:21from some sort of precarious thing
44:23like opening their oven and putting a fan next to their oven
44:27to keep their home warm, right?
44:30We've heard the stories if not done it yourselves,
44:35but I think in a warming climate,
44:37we need to start having a conversation on analogous,
44:42what I'm coining the heat stroke or go broke dilemma.
44:46What does it mean to think about that
44:49there are folks who potentially have ACs in their homes,
44:53but can't afford to run those ACs.
44:58How do we think about that
45:00they may be foregoing other important staples of their lives
45:05on the other side of things to cool their homes.
45:09And so I think that there's a real opportunity
45:12for climate epidemiology and climate and health research
45:16to engage with some of this.
45:21And finally, I'm also interested in continuing to integrate
45:25the social and environmental determinants of health.
45:29So I didn't attend the society for epidemiologic research
45:33conference this year, but I saw on Twitter
45:35that one of the big takeaways was a quote from Jay Kaufman,
45:39who said that all epidemiology is social epidemiology.
45:44And I think that that lends a real opportunity for us
45:48to think about borrowing from the social epidemiology
45:54literature and also lending our tools
45:57to the social epidemiology literature.
46:00So we recently just published a paper
46:04in Nature Communications
46:06where we actually used environmental exposure
46:10mixtures methods that were designed
46:12for the environmental health sciences,
46:15and actually implied it to thinking about neighborhood
46:18disadvantage to try and understand some of the infection
46:22disparities that we're seeing in New York city for COVID-19.
46:28And so I think that there's an opportunity here to continue
46:32to, you know, trade and learn lessons
46:35across the different areas of public health.
46:41I'm also conducting a large natality analysis that I
46:45mentioned earlier in Mexico and soon hopefully accessing
46:51data for also New York state.
46:56And we're trying to apply mixtures methods in this context
47:00as well thinking about perinatal and climate epidemiology.
47:05I also want to continue to expand
47:07my own environmental justice lens.
47:10I think a lot of focus in environmental health
47:12has been on distributive justice,
47:16but what does it mean to also think about different forms
47:19of environmental justice,
47:20like procedural justice or restorative justice
47:24in these contexts?
47:26And then finally, I'm hoping to get more engaged
47:29in community and policy engaged research to try and find
47:33climate energy and health leverage points that we can use
47:37to create a more health equitable
47:40and climate equitable future.
47:44So of course this research relies on a ton of folks to help
47:49make this possible, so thank you to all of those folks,
47:52as well as the funding that has made this all possible.
47:58And with that, I will open up for questions.
48:05<v ->So, yeah, thank you, Daniel, for a very well-presented</v>
48:09and interesting talk.
48:13I could start with a question.
48:15Well, maybe other people are thinking about theirs,
48:18so you spoke a lot about temperature exposure disparities
48:29and then introduced how energy,
48:32so you have the temperature exposure disparities,
48:36and then on top of that,
48:37you have the people with the highest temperature exposure
48:42having less of an ability to deal with that high temperature
48:46exposure and that part you didn't address as much,
48:51you know, understand that you can only do so much,
48:54but I'm wondering, you know,
48:56have you thought about ways to measure that,
49:01let's call it energy insecurity in epidemiologic studies
49:06in order to make that next step?
49:10<v ->Yeah, absolutely.</v>
49:11So I'm interested in this in two different ways.
49:15So I think that we could do work to actually collect data
49:21from folks to try and get a better sense,
49:24a better quantitative sense of people's energy insecurity.
49:31So Diana has developed actually an energy insecurity
49:37screening tool and so it would be great to try
49:42and get that screening tool out there
49:44as part of larger studies so that we can understand
49:48the kind of geographic distribution
49:51of this energy insecurity and trying to overlay that
49:55potentially with what we know about temperature.
49:59So that's on one end.
50:00On the other end, I think the lower hanging fruit
50:07is actually to access energy data.
50:10And so this is something that we're working on right now
50:14actually is to use energy data and pair that with
50:19our temperature predictions to see if we could see
50:23differences in the dose response relationship
50:28between neighborhood temperature
50:31and energy utilization by neighborhood.
50:34And if we see differences in the slopes
50:37between those neighborhoods,
50:39then that would imply to me that potentially
50:43those are differences in your response
50:46to the temperature and your ability to keep yourself cool.
50:51Of course, that needs to be adjusted
50:52for many, many different things,
50:55but that is where I'm thinking as a lower hanging fruit
51:00using administrative data at the moment.
51:04<v ->Great, other questions, comments?</v>
51:10<v ->I have a question or a comment and observation,</v>
51:13first of all, this is an amazing presentation.
51:16It's brilliant work, and it could not be more timely.
51:19And I'm going to go to your last point, talking about,
51:23you know, the application of your work and of this research
51:26within the current policy development work
51:29at the federal level right now.
51:31And I think that you're diving in and focusing in
51:34on that exposure data and how
51:36we're not getting an accurate indication of what
51:39the risk are is vitally important
51:42and there are a couple of proceedings right now, you know,
51:44with the executive order 13895,
51:48with executive order 14009.
51:50There's an OMB, a docket open until July six.
51:54There's another FEMA docket open until July 21st,
51:58is how are you, whether you are planning
52:02or whether you could consider
52:05taking your research and getting it into these
52:07and other dockets because that is setting
52:10the administrative record where we can start changing how
52:14the federal government is thinking about this.
52:17So I don't know what your thoughts are in trying to move
52:21in those spaces.
52:23<v ->Yeah, no, absolutely.</v>
52:25And I would definitely look to others who are closer
52:30to the policy landscape to help me figure out
52:34what the leverage points are.
52:36The most proximal leverage point that I'm aware of
52:41is actually what environmental justice folks
52:44are talking about right now.
52:46Folks that We Act are talking about that the low income home
52:51energy assistance program has been historically used for
52:56helping to keep folks warm during the winter,
53:01but has been lesser so used to help keep folks cool
53:06during the summer.
53:07And so we already have a policy instrument in place
53:12to identify the people who need the help,
53:15but we don't have the dollars allocated to the right part,
53:20potentially the right part of the exposure distribution.
53:25And so I think that that is the most proximal policy
53:30instrument that I'm aware of that could help move the needle
53:34towards improving public health.
53:38<v ->That's fantastic.</v>
53:40You know I would also throw out taking that
53:42as that illustration applying the national environmental
53:44policy act and the resurgence and undoing
53:47what the Trump administration did to that law
53:49because I think there's some opportunities for programmatic
53:51environmental impact statement reviews
53:54and it would be great to get your data, you know,
53:57forming the basis of some of those types of actions.
54:00So thank you.
54:01<v ->Yeah, thank you.</v>
54:04<v ->Other questions or comments?</v>
54:13<v ->Maybe just a small technical question.</v>
54:16We know that using CDC wonder data
54:18for especially the birth outcome,
54:22this issue is you mentioned briefly that the temporary
54:26resolution is not good enough.
54:28They don't accurate give you the exact date.
54:32So I'm wondering how do you deal with
54:34in your time cross data with that?
54:39<v ->Oh yeah, for sure.</v>
54:39So we ended up doing a lot of interpolation estimates.
54:47So for example CDC wonder can give you how many births
54:53there on it are in a day of the week,
54:55in a typical day of the week.
54:57And it'll give you how many births there were in a month.
55:00And so we ended up then doing a lot of averaging.
55:05Knowing Tuesdays, let's say are where, you know,
55:0930% of the births are happening,
55:1120% are happening on Wednesdays, let's say.
55:15Using that relationship, again with the longer month
55:21time span to then do a lot of smoothing and averaging
55:25to get an estimate of how many births there were.
55:28I don't think for this study
55:30we need an actual accurate number
55:33of births because at the end of the day,
55:38you're creating your truth with the simulation methods.
55:44But it's just a way of making sure that we have good
55:48representation of the different age groupings
55:52of different preterm births.
55:55Are there more 20 week olds perhaps being born in February
56:01rather than in June, right?
56:04Trying to preserve some of those distributions
56:07of the different weeks of gestation
56:12was where we spent a lot of our attention.
56:16<v ->Thanks yeah, that's makes a lot of sense.</v>
56:19And I'm more thinking of like a new addition
56:22to your similar study in the future, your future work,
56:25if you want to extend to the whole U.S.
56:27that might be something to be carefully dealt with.
56:33<v ->Yeah, absolutely.</v>
56:36<v ->So I, there's a question in the chat.</v>
56:38I think this'll be the last question.
56:40It's from Taiwo Bello,
56:42Please, how convinced are you about these studies
56:47considering that Africa has the hottest temperature
56:50and majority had no cooling systems in place
56:54and what are the limitations of your research findings?
56:57Thank you.
57:00<v ->Yeah, absolutely.</v>
57:01So I think the temperature epidemiology
57:05generally shows that there is such a thing
57:08as acclimatization, that there are differences
57:11in people's response to different temperatures
57:14in different parts of the world based on what
57:17they're historically exposed to.
57:20And so to some degree,
57:24people are climatized to the places that they live in.
57:28Another factor that needs to be considered as well is that
57:34humidity is also very different in different parts
57:37of the world.
57:38So in Western Africa, for example,
57:42at least the places that I've done research,
57:44humidity is not as high
57:47as it is in the Caribbean, let's say,
57:52or in other parts of the world, right?
57:54And so humidity plays a big part in our ability
57:57to thermo regulate in our ability to dissipate heat.
58:01And so I think that that's an important part of this
58:05relationship that a lot of temperature epidemiology
58:08kind of grapples with to do this.
58:12And I think the last thing I should mention is I think
58:18that we don't have sufficient evidence in many parts
58:21of the world to necessarily say that that heat
58:26is not an issue in Africa.
58:29There are studies that show that heat is an issue in Africa,
58:33even though the dose response relationships
58:36may be different, but nonetheless people
58:39are impacted by heat in Sub-Saharan Africa as well
58:45and I think it's actually a call for more research
58:49in the region to understand
58:52what those relationships look like.
58:57<v ->Okay, so thank you very much, Daniel.</v>
59:00You gave a very interesting talk and congratulations
59:04on doing such great work.
59:06<v ->Thank you so much.</v>
59:08<v ->Okay, take care, everyone.</v>