TheHealthEffectsofHeatWaves:PresentandFuture

RogerD.Peng,PhD

DepartmentofBiosta/s/csJohnsHopkinsBloombergSchoolofPublicHealth

WorkshoponEnvironmetrics,NCAR2010

(JointworkwithJFBobb,CTebaldi,LMcDaniel,MLBell,FDominici)

ClimateChangeandHealth

•  ClimatechangeisthoughttoaffecthealthbychangingthedistribuPonofknownriskfactors(droughts,floods,heatwaves,diseasevectors,aero‐allergens)

•  DesigningintervenPonsandmiPgaPonstrategiesrequiresunderstandingwhatpopulaPonsaremostvulnerabletoclimate‐relatedriskfactors

ClimateandHealthPathways

8.2 Current sensitivity and vulnerability

Systematic reviews of empirical studies provide the bestevidence for the relationship between health and weather orclimate factors, but such formal reviews are rare. In this section,we assess the current state of knowledge of the associationsbetween weather/climate factors and health outcome(s) for thepopulation(s) concerned, either directly or through multiplepathways, as outlined in Figure 8.1. The figure shows not onlythe pathways by which health can be affected by climate change,but also shows the concurrent direct-acting and modifying(conditioning) influences of environmental, social and health-system factors.Published evidence so far indicates that:• climate change is affecting the seasonality of some allergenicspecies (see Chapter 1) as well as the seasonal activity anddistribution of some disease vectors (see Section 8.2.8);

• climate plays an important role in the seasonal pattern ortemporal distribution of malaria, dengue, tick-borne diseases,cholera and some other diarrhoeal diseases (see Sections8.2.5 and 8.2.8);

• heatwaves and flooding can have severe and long-lastingeffects.

8.2.1 Heat and cold health effects

The effects of environmental temperature have been studiedin the context of single episodes of sustained extreme

temperatures (by definition, heatwaves and cold-waves) and aspopulation responses to the range of ambient temperatures(ecological time-series studies).

8.2.1.1 HeatwavesHot days, hot nights and heatwaves have become more

frequent (IPCC, 2007a). Heatwaves are associated with markedshort-term increases in mortality (Box 8.1). There has been moreresearch on heatwaves and health since the TAR in NorthAmerica (Basu and Samet, 2002), Europe (Koppe et al., 2004)and East Asia (Qiu et al., 2002; Ando et al., 2004; Choi et al.,2005; Kabuto et al., 2005).A variable proportion of the deaths occurring during

heatwaves are due to short-term mortality displacement (Hajatet al., 2005; Kysely, 2005). Research indicates that thisproportion depends on the severity of the heatwave and thehealth status of the population affected (Hemon and Jougla,2004; Hajat et al., 2005). The heatwave in 2003 was so severethat short-term mortality displacement contributed very little tothe total heatwave mortality (Le Tertre et al., 2006).Eighteen heatwaves were reported in India between 1980 and

1998, with a heatwave in 1988 affecting ten states and causing1,300 deaths (De and Mukhopadhyay, 1998; Mohanty andPanda, 2003; De et al., 2004). Heatwaves in Orissa, India, in1998, 1999 and 2000 caused an estimated 2,000, 91 and 29deaths, respectively (Mohanty and Panda, 2003) and heatwavesin 2003 inAndhra Pradesh, India, caused more than 3000 deaths(Government of Andhra Pradesh, 2004). Heatwaves in SouthAsia are associated with high mortality in rural populations, and

Human Health Chapter 8

396

Figure 8.1. Schematic diagram of pathways by which climate change affects health, and concurrent direct-acting and modifying (conditioning)influences of environmental, social and health-system factors.

AR4,WGII,ch.8,2007

NIHWorkingGroupReportandPAR

1

Published by Environmental Health Perspectives andthe National Institute of Environmental Health Sciences

A Human Health Perspective On Climate Change

A Report Outlining the Research Needs on the Human Health Effects of Climate Change

APRIL 22, 2010

NIHPAR‐10‐235:ClimateChangeandHealth:AssessingandModelingPopula?onVulnverabilitytoClimateChange(R21)

ChallengesinClimateChangeandHealthResearch

•  Workishighlyinterdisciplinary,requiresexperPsefromavarietyofdomains

•  Dataneedtobemerged/integratedfromavarietyofsources(neverdesignedtodothat)

•  NeedtounderstandmagnitudesofuncertainPesinthedatafromdifferentdomains

•  Notobviouswheretopublishresults!•  Fundingmechanismsonlyrecentlydeveloped

•  ButstaPsPcianscan/shouldplayabigrole!

HeatWavesandClimateChange

•  High/extremeambienttemperaturesareassociatedwithmortalityandotherhealthoutcomesintheNorthAmerica(Basu&Samet2002,Anderson&Bell2010)

•  WillclimatechangeaffectthedistribuPonofheatwaves?

HeatWavesandClimateChange

•  TherehasbeenanincreaseinthefrequencyofheatwavesinrecentPmes(“Likely”,IPCC,2007)

•  Thefrequencyandseverityofheatwaveswillincreaseinthefuture(“VeryLikely”,IPCC,2007;Meehl&Tebaldi2004)

Howwillanychangeinthedistribu/onofheatwavesaffectmortalityandmorbidity?

MoreIntense,FrequentHeatWaves

Meehl&Tebaldi2004

25%(Chicago)and31%(Paris)increaseinheatwavefrequencyin2080—2099

shift in the model to more and longer livedheat waves in future climate.

Heat waves are generally associated withspecific atmospheric circulation patterns repre-sented by semistationary 500-hPa positiveheight anomalies that dynamically produce sub-sidence, clear skies, light winds, warm-air ad-vection, and prolonged hot conditions at thesurface (15, 19). This was the case for the 1995Chicago heat wave and 2003 Paris heat wave(Fig. 3, A and B), for which 500-hPa heightanomalies of over !120 geopotential meters(gpm) over Lake Michigan for 13 to 14 July1995 and !180 gpm over northern France for 1to 13 August 2003 are significant at greater thanthe 5% level according to a Student’s t test. Astratification based on composite present-dayheat waves from the model for these twolocations over the period of 1961 to 1990(Fig. 3, C and D) shows comparable ampli-tudes and patterns, with positive 500-hPaheight anomalies in both regions greater than!120 gpm and significance exceeding the5% level for anomalies of that magnitude.

There is an amplification of the positive500-hPa height anomalies associated with agiven heat wave for Chicago and Paris forfuture minus present climate (Fig. 4, A andB). Statistically significant (at greater thanthe 5% level) ensemble mean heat wave 500-hPa differences for Chicago and Paris in thefuture climate compared with present-day arelarger by about 20 gpm in the model (com-paring Fig. 4, A and B, with Fig. 3, C and D).

The future modification of heat wave char-acteristics with a distinct geographical pattern(Fig. 1, E and F) suggests that a change inclimate base state from increasing greenhousegases could influence the pattern of those chang-es. The mean base state change for future climateshows 500-hPa height anomalies of nearly !55gpm over the upper Midwest, and about !50gpm over France for the end of the 21st century(Fig. 4, C and D, all significant above the 5%level). The 500-hPa height increases over theMediterranean and western and southern UnitedStates for future climate are directly associatedwith more intense heat waves in those regions(Fig. 1, E and F), thus confirming the link be-tween the pattern of increased 500-hPa heightsfor future minus present-day climate and in-creased heat wave intensity in the future climate.A comparable pattern is present in an ensembleof seven additional models for North Americafor future minus present-day climate, with some-what less agreement over Europe (fig. S1). Inthat region, there is still the general character oflargest positive anomalies over the Mediterra-nean and southern Europe regions, and smallerpositive anomalies to the north (fig. S1), butlargest positive values occur near Spain, as op-posed to the region near Greece as in our model(Fig. 4D). This also corresponds to a similarpattern for increased standard deviations of bothsummertime nighttime minimum and daytime

Fig. 2. Based on the threshold definition of heat wave (16), mean number of heat waves per yearnear Chicago (A) and Paris (B) and mean heat wave duration near Chicago (C) and Paris (D) areshown. In each panel, the blue diamond marked NCEP indicates the value computed fromNCEP/NCAR reanalysis data. The black segment indicates the range of values obtained from thefour ensemble members of the present-day (1961 to 1990) model simulation. The red segmentindicates the range of values obtained from the five ensemble members of the future (2080 to2099) model simulation. The single members are marked by individual symbols along the segments.Dotted vertical lines facilitate comparisons of the simulated ranges/observed value.

-50

0

50

100

150

Observed Heat Wave 500hPa Height Anomalies July 13-14, 1995, minus July 1948-2003

0

50

100

150

200

Observed Heat Wave 500hPa Height Anomalies August 1-13, 2003, minus August 1948-2003

-50

0

50

100

150

Simulated Composite Heat Wave 500 hPa Height Anomalies (JJA, 1961-1990)

0

50

100

150

200

Simulated Composite Heat Wave 500 hPa Height Anomalies (JJA, 1961-1990)

48.8°N

48.8°N

23.7°N

23.7°N

65.5°N

65.5°N

26.5°N

26.5°N

123.7°W

123.7°W

70.3°W

70.3°W

16.8°W

16.8°W

30.9°E

30.9°E

A B

C D

Fig. 3. Height anomalies at 500 hPa (gpm) for the 1995 Chicago heat wave (anomalies for 13 to 14 July1995 from July 1948 to 2003 as base period), from NCEP/NCAR reanalysis data (A) and the 2003 Parisheat wave (anomalies for 1 to 13 August 2003 from August 1948 to 2003 as base period), fromNCEP/NCAR reanalysis data (B). Also shown are anomalies for events that satisfy the heat wave criteriain the model in present-day climate (1961 to 1990), computed at grid points near Chicago (C) and Paris(D). In both cases, the base period is summer [ June, July, August (JJA)], 1961 to 1990.

R E P O R T S

13 AUGUST 2004 VOL 305 SCIENCE www.sciencemag.org996

on

De

ce

mb

er

4,

20

08

w

ww

.scie

nce

ma

g.o

rgD

ow

nlo

ad

ed

fro

m

Chicago1995HeatWave

•  AmajorheatwaveoccurredinChicagoinJuly,1995

•  >700excessdeathswereajributedtotheheatwaveina1‐weekperiod

•  Elderlyandhome‐boundweremostsuscepPble•  TheheatwaveaffectedsurroundingmidwesternciPesaswell

•  AsecondmajorheatwaveoccurredinJuly1999withlessdevastaPngconsequences

Whitmanetal.1997

Chicago1999HeatWave

0

10

20

30

40

50

(a)

He

at

str

oke

ad

mis

sio

ns

May 29 Jun 18 Jul 08 Jul 28 Aug 17

60

65

70

75

80

85

(b)

Te

mp

era

ture

(°°F

)

May 29 Jun 18 Jul 08 Jul 28 Aug 17

Figure 3: (a) Numbers of emergency hospital admissions for heat stroke among Medicare enrollees inChicago, IL, June–August, 1999; (b) Daily 24-hour average temperature, June–August, 1999. Light graylines indicate values in years 2000–2006.

assumes that the time series of the outcome of interest Yt is Poisson distributed with mean µt where

log µct = !c + "c xc

t + #c zct + $c xc

t · zct + log N c

t + s(t, d)

The time series xct is the indicator of heat wave days (xc

t = 1 for a heat wave day) and zct is the daily time

series of an air pollutant. We can include multiple pollutants if necessary, in which case we will have theterms #c

1 zc1t, #c

2 zc2t, etc. By fitting the model above, we can examine estimates of $c to determine if heat

waves have a synergistic relationship with ambient air pollution exposure. The estimates of $c can becombined using the Bayesian hierarchical models described earlier to borrow strength from neighboringlocations and obtain more precise esitmates.

3.2 Predicting Future Health Impacts of Heat Waves

For evaluating the health impacts of heat waves in the future, we will link our risk estimates with dataobtained from the WCRP CMIP3 Multi-Model Dataset [12]. This dataset contains multi-model ensemblesof global forecasts from coupled atmosphere-ocean general circulation models run for a present daycontrol experiment and for 21st century experiments with various plausible forcing parameters, includingranges of CO2 concentration. For each scenario, the multi-model ensemble represents a set of possibleoutcomes, that can ideally be combined to represent the best estimate for the probability distribution of thefuture climate. This dataset contains several variables pertinent to our heat wave model, including indicesof heat wave events and daily mean/max estimates of temperature and humidity to which we could applythe model of heat waves developed in Aim 1. The availability of model predictions of these parameterswill allow us to explore the range of possible future health impacts. While there is substantial uncertaintyregarding the nature of future emissions scenarios and climate-related interventions, we will use theprobability distribution of heat waves estimated from the ensemble of forecasts to provide quantitativeupper and lower bounds on the future health impacts of heat waves.

As a product of Aim 1, we will obtain estimates of "c, the relative risk between heat waves and mortal-ity/morbidity, for different US locations. We can calculate the average number of excess deaths (ED) on aheat wave day as N c ! ("c " 1) where N c is the average number of daily deaths on a non-heat wave dayin location c. Then we can compute the annual average number of excess deaths per year by multiplying

8

European2003HeatWave

•  August4—13,2003•  EsPmated14,800excessdeathsinFranceduringthis9dayperiod

•  Temperaturesover99oFfor9consecuPvedays

•  Minimumtemperaturewas78oFduringthatperiod

•  ExcessmortalityalsoseeninItaly,Spain,Portugal,UK

Bouchama2004

WhatisaHeatWave?

•  ThereisnogenerallyagreedupondefiniPonofaheatwave

•  ExceedancesofpercenPlesofthetemperaturedistribuPon

•  Exceedancesofspecificabsolutetemperaturelevels

•  ConPnuousstretchesofhightemperature

•  Highhumidity

Present‐dayextremeheatrisk

ProjecPon1

ProjecPon2

FutureExtremeHeatExcessMortality/Morbidity(2081—2100)

DatabasesforcurrentcondiPons

Globalclimatemodels(IPCCCMIP3)

.

.

.ProjecPonN

PopulaPongrowth,agestructure(IIASA)

Weather(NCDC)

AirPolluPon(EPA)

EnvironmentalcondiPons

LocalprojecPon1

LocalprojecPon2

LocalprojecPonN

.

.

SpaPaldownscaling

FuturecondiPons

Futureextremeheatepisodes

Baselinemortality/morbidityrate

AdaptaPon

.

.Mortality(1987—2005)Morbidity(1999—2009)

NCHSandMedicare

Biological,environmental,socio‐economicmodifiersofvulnerability(Census)

ClimateScenario

MortalityData

•  ObtainedfromtheNaPonalCenterforHealthStaPsPcsfortheperiod1987—2005

•  DeathcerPficatedatacontaininganindividual’sdateandcauseofdeath,locaPonofresidence,age

•  WeconstructadailyPmeseriesofmortalityfornon‐accidentalcausesanddifferentagecategories

•  Dataareavailablefor>100metropolitanareas

Present‐dayTemperatureData

•  HourlytemperatureanddewpointtemperatureavailablefromNCDC

•  Maximumovera24hourperiodisused

•  MaximumovermulPplemonitorvaluesinacityistaken,ifavailable

NMMAPS data: 105 Cities, 1987-2005

Y ct : Number of deaths from all causes on day t in city c,

stratified by three age categories

temperaturect

confoundersct

7 / 25

NCHS105CiPes,1987‐2005

ChicagoMortalityData,1987—2005

HeatWaveDefiniPon

•  Threshold1(T1)isthe97.5thpercenPleofthedistribuPonofdailymaximumtemperatures

•  Threshold2(T2)isthe81stpercenPleofdailymaxtemperatures

•  AheatwaveisdefinedasthelongestperiodofconsecuPvedayssaPsfying:– DailymaxtempisaboveT1for>3days– DailymaxtempisaboveT2fortheenPreperiod– Averageofdailymaxtempovertheperiodis>T1

Huthetal.2000;Meehl&Tebaldi2004

ChicagoMaxTemperatureApril—September,1987—2005

StaPsPcalModelforPresent‐dayRisk

where we allow Yt to be a member of the quasipoisson family and f (·) and g(·) are unknown

smooth functions. Here, weather variables may include one or more of the following covari-

ates: current day’s maximum temperature, average of previous three days’ maximum daily

temperature, and current day’s dew point temperature. The potential confounding variables we

accounted for were current day’s mean ozone and particulate matter levels and smooth tempo-

ral fluctuations in time. The smooth function of time was included in the model to remove any

medium- and long-term trends in the data because we are only interested in examining short-

term effects of heat waves. We also stratified our analysis by three age categories (< 65 years

of age, 65–74, and ! 75) and therefore included an intercept for age and interactions of the

weather variables with age category, in the model. Interactions with age categories are needed

because of the differing temporal trends in mortality by age category. The final model was of

the form

logE[Yt ] = !1 +3

!i=2

!iI(aget = i)+3

!i=1

fi(weathert)I(aget = i)+g(confounderst) (1)

We fit several models of the form of (1), where the models differed based on which weather

covariates were included. The models were fit with the gam() function in the R package mgcv.

The final exposure-response model was selected based on the generalized cross validataion

(GCV) criterion. The next step was to apply this full model of the weather-mortality relationship

to the half of the year containing the summer season to estimate the relative risk of mortality

comparing heat wave days to non-heat wave days in Chicago for the period 1987–2005. Using

quasi-likelihood procedures (7), we obtain f̂i, the estimate of the exposure-response function for

weather and mortality. We also obtain an estimate of g, but because g is a nuisance parameter

its specific value is not of primary interest.

Relative risks were calculated separately for each age category as well as a pooled relative

risk across the three age groups. For the ith age group, the expected mortality on day t given

5

Agecategory‐specificdailymortalitymodel

SmoothsplinefuncPonoftemperaturevariables

Expectedagecategory‐specificmortalitycountondayt

SmoothfuncPonsofairpolluPonlevels(ozone,PM),temporaltrends

RangeofModelsModels f (temperature;!) df of natural

cubic splines1 !1tmax

2 !1tmax(3)

3 !1dptp

4 !1tmax+ !2tmax(3)

5 !1tmax+ !2dptp

6 !1tmax(3) + !2dptp

7 !1tmax+ !2tmax(3) + !3dptp

8–12 ns(tmax;!,") " ! {2, . . . , 6}13–17 ns(tmax(3);!,") " ! {2, . . . , 6}18–22 ns(dptp;!,") " ! {2, . . . , 6}23–27 ns(tmax;!,") + ns(tmax(3);!,") " ! {2, . . . , 6}

28 ns(tmax;!,") + ns(dptp;!,") " = 329 ns(tmax(3);!,") + ns(dptp;!,") " = 330 ns(tmax;!,") + ns(tmax(3);!,") + ns(dptp;!,") " = 331 ns(tmax;!,")" ns(tmax(3);!,") " = 332 ns(tmax;!,")" ns(dptp;!,") " = 333 ns(tmax(3);!,")" ns(dptp;!,") " = 3

12 / 25

RelaPveRiskEsPmate

the weather variables and confounders on that day is

E[Yt | weathert ,confounderst ] = exp{!i + fi(weathert)+g(confounderst)}

It follows that the relative risk associated with a heat wave day for age group i is given by

RRi =E[Yt | hwt = 1,confounderst ]E[Yt | hwt = 0,confounderst ]

We estimate this quantity by

!RRi =1n1

!t exp{ f̂i(weathert)}I(hwt = 1)1n0

!t exp{ f̂i(weathert)}I(hwt = 0),

where n1 is the number of heat wave days and n0 is the number of non-heat wave days in

Chicago during this period and I(·) is an indicator function. Similarly, the relative risk pooled

across age categories is estimated by

!RR =1n1

!t !i exp{ f̂i(weathert)}I(hwt = 1)1n0

!t !i exp{ f̂i(weathert)}I(hwt = 0)

We calculate variances and asympototic 95% confidence intervals for the relative risk estimates

by applying the delta method (8).

The expected number of excess deaths on a heat wave day compared to a non-heat wave day

was calculated as EDhw = N! (RR" 1) where N is the number of daily deaths on a non-heat

wave day averaged across the study period and RR is the relative risk of mortality associated

with heat wave days, i.e. the ratio of the rate of mortality on a heat wave day and the rate of mor-

tality on a non-heat wave day. Annual excess heat wave deaths were obtained by multiplying

EDhw by the average number of heat waves days per year.

In the second stage of our approach we obtained estimates of future heat waves from 11

different climate model simulations of temperature from the Program for Climate Model Di-

agnosis and Intercomparison (PCMDI) as part of the Coupled Model Intercomparison Project

(CMIP3) (9). The frequency and length of heat waves for the 2081–2100 period was estimated

6

the weather variables and confounders on that day is

E[Yt | weathert ,confounderst ] = exp{!i + fi(weathert)+g(confounderst)}

It follows that the relative risk associated with a heat wave day for age group i is given by

RRi =E[Yt | hwt = 1,confounderst ]E[Yt | hwt = 0,confounderst ]

We estimate this quantity by

!RRi =1n1

!t exp{ f̂i(weathert)}I(hwt = 1)1n0

!t exp{ f̂i(weathert)}I(hwt = 0),

where n1 is the number of heat wave days and n0 is the number of non-heat wave days in

Chicago during this period and I(·) is an indicator function. Similarly, the relative risk pooled

across age categories is estimated by

!RR =1n1

!t !i exp{ f̂i(weathert)}I(hwt = 1)1n0

!t !i exp{ f̂i(weathert)}I(hwt = 0)

We calculate variances and asympototic 95% confidence intervals for the relative risk estimates

by applying the delta method (8).

The expected number of excess deaths on a heat wave day compared to a non-heat wave day

was calculated as EDhw = N! (RR" 1) where N is the number of daily deaths on a non-heat

wave day averaged across the study period and RR is the relative risk of mortality associated

with heat wave days, i.e. the ratio of the rate of mortality on a heat wave day and the rate of mor-

tality on a non-heat wave day. Annual excess heat wave deaths were obtained by multiplying

EDhw by the average number of heat waves days per year.

In the second stage of our approach we obtained estimates of future heat waves from 11

different climate model simulations of temperature from the Program for Climate Model Di-

agnosis and Intercomparison (PCMDI) as part of the Coupled Model Intercomparison Project

(CMIP3) (9). The frequency and length of heat waves for the 2081–2100 period was estimated

6

the weather variables and confounders on that day is

E[Yt | weathert ,confounderst ] = exp{!i + fi(weathert)+g(confounderst)}

It follows that the relative risk associated with a heat wave day for age group i is given by

RRi =E[Yt | hwt = 1,confounderst ]E[Yt | hwt = 0,confounderst ]

We estimate this quantity by

!RRi =1n1

!t exp{ f̂i(weathert)}I(hwt = 1)1n0

!t exp{ f̂i(weathert)}I(hwt = 0),

where n1 is the number of heat wave days and n0 is the number of non-heat wave days in

Chicago during this period and I(·) is an indicator function. Similarly, the relative risk pooled

across age categories is estimated by

!RR =1n1

!t !i exp{ f̂i(weathert)}I(hwt = 1)1n0

!t !i exp{ f̂i(weathert)}I(hwt = 0)

We calculate variances and asympototic 95% confidence intervals for the relative risk estimates

by applying the delta method (8).

The expected number of excess deaths on a heat wave day compared to a non-heat wave day

was calculated as EDhw = N! (RR" 1) where N is the number of daily deaths on a non-heat

wave day averaged across the study period and RR is the relative risk of mortality associated

with heat wave days, i.e. the ratio of the rate of mortality on a heat wave day and the rate of mor-

tality on a non-heat wave day. Annual excess heat wave deaths were obtained by multiplying

EDhw by the average number of heat waves days per year.

In the second stage of our approach we obtained estimates of future heat waves from 11

different climate model simulations of temperature from the Program for Climate Model Di-

agnosis and Intercomparison (PCMDI) as part of the Coupled Model Intercomparison Project

(CMIP3) (9). The frequency and length of heat waves for the 2081–2100 period was estimated

6

RelaPveriskforindividualagecategories

OverallrelaPveriskacrossallagecategories

(Varianceobtainedviadeltamethod)

PriorDistribuPonsPrior selection

Adopted class of prior distributions for GLMs from Raftery (1996)

!(!k ,"k | Mk )

Accounts for nesting of models, e.g.

!(("0,"1) | "1 = 0) = !("0)

Depends on three hyperparameters #, $, and %

Only # has significant impact on inference !" reportinferences across a range of values of #

Var(RR | Mk ) # #2&k

15 / 25

PosteriorDistribuPonsforLogRRResults

Atlanta

!

-0.02 0.00 0.02 0.04 0.06 0.08

010

20

30

Chicago

!

0.09 0.10 0.11 0.12 0.13 0.14

020

40

60

Cleveland

!

-0.02 0.00 0.02 0.04 0.06 0.08

05

15

25

Detroit

!

0.06 0.08 0.10 0.12

010

20

30

Dallas/Fort Worth

!

-0.01 0.00 0.01 0.02 0.03

020

40

60

Houston

!

-0.02 -0.01 0.00 0.01 0.02

0100

200

Los Angeles

!

0.020 0.030 0.040 0.050

020

40

Miami

!

-0.01 0.00 0.01 0.02 0.030

40

80

Minneapolis/St. Paul

!

-0.02 0.00 0.02 0.04

020

40

New York

!

0.080 0.090 0.100 0.110

020

40

60

80

Oakland

!

0.02 0.04 0.06 0.08 0.10 0.12

05

15

25

Philadelphia

!

0.02 0.04 0.06 0.08 0.10

010

20

30

40

Phoenix

!

0.00 0.02 0.04 0.06

010203040

Riverside

!

0.00 0.02 0.04 0.06

010

20

30

40

San Antonio

!

-0.04 -0.02 0.00 0.02 0.04

010

30

San Bernardino

!

0.00 0.02 0.04 0.06 0.08 0.10

010

20

30

San Diego

!

0.00 0.01 0.02 0.03 0.04 0.05 0.06

020

40

60

San Jose

!

0.05 0.06 0.07 0.08 0.09 0.10

020

40

Seattle

!

0.03 0.04 0.05 0.06 0.07 0.08

010203040

Santa Ana/Anaheim

!

-0.01 0.00 0.01 0.02 0.03 0.04

020

40

Figure: Histograms and kernel density estimates of 2000 samples fromthe posterior P(!c | yc) for the 20 largest cities, where ! is the logrelative risk of mortality comparing heat wave to non-heat wave days.

19 / 25

PosteriorDistribuPonsforLogRR(105CiPes,1987—2005)Results

Figure: 95% highest posterior density intervals for the log relative riskof mortality. Cities categorized into 7 regions: southeast (SE),southwest (SW), southern California (SC), northeast (NE), uppermidwest (UM), industrial midwest (IM), and northwest (NW). Withinregions, cities listed from left to right in order of decreasing latitude.

20 / 25

PosteriorModeforLogRR(105CiPes,1987—2005)

Results

<20th percentile

20th-40th percentile

40th-60th percentile

60th-80th percentile

>80th percentile

Figure: Estimated % increase in mortality associated with a heat waveday (posterior mode). The 20th, 40th, 60th, and 80th percentiles are0.2%, 1.1%, 2.2%, 4.4%, and 12.4%, respectively. Size ! !̂!1

BMA21 / 25

FutureTemperature

•  GCMoutputobtainedfromtheWCRPCMIP3mulP‐modeldatasetarchiveatPCMDI

•  Dailymax(surface)temperaturefrom7modelswereobtainedfor1981—2000and2081—2100

•  EmissionsscenariosA1B,B1,andA2wereused

•  Numberofheatwavesandlengthofheatwavescomputedforeachmodel

ClimateScenarios

IPCC,SRES,2000

Emissions Scenarios4

The main characteristics of the four SRES storylines and scenario families

By 2100 the world will have changed in ways that are difficult to imagine – as difficult as it would have been at the end of the

19th century to imagine the changes of the 100 years since. Each storyline assumes a distinctly different direction for future

developments, such that the four storylines differ in increasingly irreversible ways. Together they describe divergent futures that

encompass a significant portion of the underlying uncertainties in the main driving forces. They cover a wide range of key

“future” characteristics such as demographic change, economic development, and technological change. For this reason, their

plausibility or feasibility should not be considered solely on the basis of an extrapolation of current economic, technological,

and social trends.

• The A1 storyline and scenario family describes a future world of very rapid economic growth, global population that

peaks in mid-century and declines thereafter, and the rapid introduction of new and more efficient technologies. Major

underlying themes are convergence among regions, capacity building, and increased cultural and social interactions, with

a substantial reduction in regional differences in per capita income. The A1 scenario family develops into three groups

that describe alternative directions of technological change in the energy system. The three A1 groups are distinguished

by their technological emphasis: fossil intensive (A1FI), non-fossil energy sources (A1T), or a balance across all sources

(A1B).3

Figure 1: Schematic illustration of SRES scenarios. Four qualitative storylines yield four sets of scenarios called “families”:

A1, A2, B1, and B2. Altogether 40 SRES scenarios have been developed by six modeling teams. All are equally valid with

no assigned probabilities of occurrence. The set of scenarios consists of six scenario groups drawn from the four families:

one group each in A2, B1, B2, and three groups within the A1 family, characterizing alternative developments of energy

technologies: A1FI (fossil fuel intensive), A1B (balanced), and A1T (predominantly non-fossil fuel). Within each family and

group of scenarios, some share “harmonized” assumptions on global population, gross world product, and final energy.

These are marked as “HS” for harmonized scenarios. “OS” denotes scenarios that explore uncertainties in driving forces

beyond those of the harmonized scenarios. The number of scenarios developed within each category is shown. For each of

the six scenario groups an illustrative scenario (which is always harmonized) is provided. Four illustrative marker scenarios,

one for each scenario family, were used in draft form in the 1998 SRES open process and are included in revised form in

this Report. Two additional illustrative scenarios for the groups A1FI and A1T are also provided and complete a set of six

that illustrates all scenario groups. All are equally sound.

3 Balanced is defined as not relying too heavily on one particular energy source, on the assumption that similar improvement rates apply

to all energy supply and end use technologies.

IPCCA1Family

“TheA1storylineandscenariofamilydescribesafutureworldofveryrapideconomicgrowth,globalpopula?onthatpeaksinmid‐centuryanddeclinesthereaqer,andtherapidintroducPonofnewandmoreefficienttechnologies.Majorunderlyingthemesareconvergenceamongregions,capacitybuilding,andincreasedculturalandsocialinteracPons,withasubstanPalreducPoninregionaldifferencesinpercapitaincome.”

“TheA1scenariofamilydevelopsintothreegroupsthatdescribealternaPvedirecPonsoftechnologicalchangeintheenergysystem.ThethreeA1groupsaredisPnguishedbytheirtechnologicalemphasis:fossilintensive(A1FI),non‐fossilenergysources(A1T),orabalanceacrossallsources(A1B)”

IPCC,SRES,2000

PopulaPonTrends

IPCC,SRES,2000

ChangingAgeStructureofPopulaPon

IPCC,SRES,2000

EsPmateofRRforagecategory>65was~30%largerthanoverallRR

ChicagoGridCell

!

GFDLCM2.0

HeatWaveMortalityEsPmate

the weather variables and confounders on that day is

E[Yt | weathert ,confounderst ] = exp{!i + fi(weathert)+g(confounderst)}

It follows that the relative risk associated with a heat wave day for age group i is given by

RRi =E[Yt | hwt = 1,confounderst ]E[Yt | hwt = 0,confounderst ]

We estimate this quantity by

!RRi =1n1

!t exp{ f̂i(weathert)}I(hwt = 1)1n0

!t exp{ f̂i(weathert)}I(hwt = 0),

where n1 is the number of heat wave days and n0 is the number of non-heat wave days in

Chicago during this period and I(·) is an indicator function. Similarly, the relative risk pooled

across age categories is estimated by

!RR =1n1

!t !i exp{ f̂i(weathert)}I(hwt = 1)1n0

!t !i exp{ f̂i(weathert)}I(hwt = 0)

We calculate variances and asympototic 95% confidence intervals for the relative risk estimates

by applying the delta method (8).

The expected number of excess deaths on a heat wave day compared to a non-heat wave day

was calculated as EDhw = N! (RR" 1) where N is the number of daily deaths on a non-heat

wave day averaged across the study period and RR is the relative risk of mortality associated

with heat wave days, i.e. the ratio of the rate of mortality on a heat wave day and the rate of mor-

tality on a non-heat wave day. Annual excess heat wave deaths were obtained by multiplying

EDhw by the average number of heat waves days per year.

In the second stage of our approach we obtained estimates of future heat waves from 11

different climate model simulations of temperature from the Program for Climate Model Di-

agnosis and Intercomparison (PCMDI) as part of the Coupled Model Intercomparison Project

(CMIP3) (9). The frequency and length of heat waves for the 2081–2100 period was estimated

6

Excessdeathsonaheatwaveday

Baseline#ofdeathsonanon‐heatwaveday(esPmatedfrompresent‐daydata)

RelaPveriskassociatedwithaheatwaveday

AssumpPonsforFutureEsPmates

•  ConstantheatwaverelaPveriskoverPme•  NoadaptaPontoextremeheat

•  Constantrateofmortalityonnon‐heatwavedays

•  Minimalshort‐termmortalitydisplacement

AnnualHeatWaveFrequencyandAverageHeatWaveLengthforChicago,2081—2100

A Tables

Table 1: Estimates of the number of heat waves per year (annual frequency), average heat wave

length (in days), and excess deaths per year from heat waves days for Chicago, 2081–2100.

Heat Waves Excess DeathsFrequency Length # per year 95% CI

Model (# per year) (days)

LASG/IAP FGOALS-g1.0 1.8 9.8 165 (130, 200)CCCMA CGCM 3.1 (T63) 1.4 12.5 170 (134, 206)CCCMA CGCM 3.1(T47) 1.4 13.1 172 (135, 208)CSIRO Mk 3.0 1.4 14.4 196 (154, 237)GFDL CM 2.0 2.1 14.6 286 (225, 348)GISS AOM 2.0 16.0 300 (236, 364)MRI CGCM 2.3.2a 2.5 16.4 392 (309, 477)CNRM CM3 3.3 25.9 800 (629, 971)MPI ECHAM5 5.2 17.6 855 (673, 1038)MIROC 3.2 (medres) 5.4 19.6 990 (779, 1203)MIROC 3.2 (hires) 6.3 22.8 1345 (1059, 1634)

11

ForChicago1987—2005therewere0.7heatwavesperyearwithanaveragelengthof9.2days

AnnualHeatWaveMortalityforChicago,2081—2100

!

!

!

!

!

!

0 500 1000 1500 2000 2500 3000

Annual heat wave mortality

SR

ES

Scenari

o

B1

A1B

A2

!

!

csiro.mk3.0

cccma.cgcm3.1

gfdl.cm2.0

mri.cgcm2.3.2a

mpi.echam5

cnrm.cm3

miroc3.2.medres

1995 heat !

wave!

1999 heat!

wave!

Summary

•  Using19yearsofdailydatafor1987—2005,wefoundstrongevidencethatheatwavesareassociatedwithexcessmortalityinpresent‐dayChicago

•  TheincidenceandlengthofheatwavesinChicagoisprojectedtoincreasein2081—2100;therewasawiderangeofvariaPonacrossacollecPonof7GCMs

•  MortalityfromheatwavesisesPmatedtoincreasefrompresent‐daycondiPonsbyafactorrangingfrom3to25dependingontheGCMused

FutureDirecPons

•  MethodologyisreproducibleacrossdifferentlocaPonsusingpubliclyavailabledata

•  LooknaPonallyatvariaPoninheatwaveriskacrosslocaPonsviahierarchicalmodeling

•  Lookatmorbidityeffectsofheat•  IncorporatemorerealisPcassumpPonsaboutthefuture(adaptaPon)

•  ExaminePme‐varyingrisk,relatetootherfactors•  DistributedlagmayesPmatethe“totaleffect”ofaheatwavebejerthansinglelagmodels(mortalitydisplacement/harvesPng)

Collaborators

•  JenniferF.Bobb(JHSPH)•  FrancescaDominici(Harvard)

•  ClaudiaTebaldi(ClimateCentral/UBC)

•  MichelleL.Bell(Yale)

•  LarryMcDaniel(NCAR)

JABESSpecialIssueonClimateChangeandHealth

•  Co‐Editors:RogerPeng,BoLi•  Submissiondeadline:~October2011

•  Lookingforpaperson– Climatemodeling– Modelingofriskfactors/environmentalexposures– Healtheffectsofclimate‐relatedphenomena

– Predictorsofvulnerabilitytoclimatechange

SensiPvityAnalysisSensitivity Analysis

0.00 0.02 0.04 0.06

010203040

Atlanta

!

1

1.65

3

5

BIC

0.09 0.10 0.11 0.12 0.13 0.14

020

40

60

Chicago

!

-0.02 0.00 0.02 0.04 0.06 0.08

010

30

50

Cleveland

!

0.06 0.08 0.10 0.12

020

40

60

Detroit

!

-0.01 0.00 0.01 0.02 0.03

020

40

60

Dallas/Fort Worth

!

-0.015 -0.005 0.005 0.015

0100

300

Houston

!

0.020 0.030 0.040 0.050

040

80

Los Angeles

!

0.000 0.010 0.020

0100

200

Miami

!

-0.02 0.00 0.02 0.04

020

40

60

Minneapolis/St. Paul

!

0.085 0.095 0.105 0.115

020

60

New York

!

0.02 0.04 0.06 0.08 0.10 0.12

010

20

30

Oakland

!

0.04 0.06 0.08 0.10

010

30

Philadelphia

!

0.00 0.02 0.04 0.06

020

40

Phoenix

!

0.00 0.02 0.04 0.06

010203040

Riverside

!

-0.02 0.00 0.02 0.04

010

30

50

San Antonio

!

0.02 0.04 0.06 0.08 0.10

010

20

30

San Bernardino

!

0.00 0.01 0.02 0.03 0.04 0.05

020406080

San Diego

!

0.05 0.06 0.07 0.08 0.09 0.10

020

40

San Jose

!

0.02 0.04 0.06 0.08

010

30

50

Seattle

!

-0.01 0.00 0.01 0.02 0.03 0.04

020

40

60

Santa Ana/Anaheim

!

Figure: Kernel density estimates of the posterior P(! | y) under BMAfor 4 values of hyperparameter ", and of the posterior under theBIC-selected model P(! | MBIC , y).

24 / 25

Time‐varyingRelaPveRisk

1980 to 2001 show that air conditioning use has been steadilyincreasing in all areas of the United States.22 Although anincrease in air conditioning is a plausible explanation of thedecrease in heat-related cardiovascular deaths, it is con-founded with other changes over time, such as improvedhealth care.

Cold-Related MortalityWhile an increase in air conditioning over time may

have affected heat-related cardiovascular deaths, nothing haschanged the impact of cold temperatures on mortality. Themechanism by which cold temperatures lead to increasedcardiovascular deaths is most likely via blood pressure.23

Susceptibility to cold-related mortality has been associated

with race,24,25 education24 and female sex.3,25 The sex differ-ence suggests either that clothing is an important modifier26 orthat there is a biologic difference between the ability tothermoregulate. Body temperature is regulated by the hypo-thalamus neurons, which are directly influenced by estrogenthrough estrogen receptors. The associations with race andeducation suggest a socioeconomic effect, although results onthe socioeconomic effect on cold-related deaths have beenmixed. Of 2 large UK studies, one found an associationbetween cold homes and increased risk of death,10 but an-other found that deprivation in an area was not related to riskof excess winter all-cause mortality.27

It is plausible that improvements over time in thestandard of living (specifically housing quality and heating)would reduce the number of cold-related deaths. The resultsfrom this study suggest either that improvements in the USstandard of living were insufficient or that such improve-ments are in fact not protective. Another possible pathway toprotection of the elderly from low temperatures is more andbetter clothes in cold weather.26 The results shown heresuggest that protective measures need to be taken not just inwinter, but also in relatively cold days spring and fall.However, there is no direct evidence in the literature tosupport an intervention of increased clothing. Basu andSamet1 have provided guidelines for the future research intoheat-related mortality. The results from the present studyindicate that new studies of cold-related cardiovasculardeaths are also needed. To date most research has analyzedtemperature at a population level, using temperature measure-ments obtained from outdoor monitors. Although logisticallymore difficult, a cohort study that monitored temperature insubjects’ homes and collected details on subjects’ clothingwould have much greater power to detect differences in riskrelated to individual and socioeconomic factors.

A successful intervention for cold-related mortalitycould have a substantial public health impact. Using the data

FIGURE 2. Mean changes in daily cardiovascular deaths (%)due to a 10°F increase temperature in summer and winter byregion.

FIGURE 1. Mean changes in daily cardiovasculardeaths (%) and 95% posterior intervals due to a10°F increase in temperature by year and season.

Epidemiology • Volume 18, Number 3, May 2007 Temperature and Cardiovascular Deaths in the US Elderly

© 2007 Lippincott Williams & Wilkins 371

Barnej,2007