TIME SERIES ANALYSIS FOR STUDIES OF WEATHER AND HEALTH Paul Wilkinson Public & Environmental Health Research Unit London School of Hygiene & Tropical Medicine Keppel Street London WC1E 7HT (UK)
Cardiovascular deaths/day 01jan199001jan199101jan199201jan199301jan1994 CVD deaths Mean temperature LONDON,
CLIMATE OR WEATHER?
TWO APPROACHES Episode analysis - transparent - risk defined by comparison to local baseline Regression analysis of all days of year - uses full data set - requires fuller data and analysis of confounders - can be combined with episode analysis
EPISODES
No. of deaths/day Date Influenza ‘epidemic’ Period of heat Smooth function of date with control for influenza Smooth function of date Triangle: attributable deaths PRINCIPLES OF EPISODE ANALYSIS
INTERPRETATION Common sense, transparent Relevant to PH warning systems But How to define episode? - relative or absolute threshold - duration - composite variables Uses only selected part of data Most sophisticated analysis requires same methods as for regression of all days of year
REGRESSION OF ALL DAYS
TIME-SERIES Short-term temporal associations Usually day to day fluctuations over several years Similar to any regression analysis but with specific features Methodologically sound (same population compared with itself day by day)
Time-varying confounders influenza day of the week, public holidays pollution Secular trend Season STATISTICAL ISSUES 1
STATISTICAL ISSUES 1I Shape of exposure-response function smooth functions linear splines Lags simple lags distributed lags Temporal auto-correlation
Source: Anderson HR, et al. Air pollution and daily mortality in London: Br Med J 1996; 312:665-9
TEMPERATURE DEPENDENCE OF DAILY MORTALITY, LONDON
THE MODEL… (log) rate =ß 0 + ß 1 (high temp.)+ ß 2 (low temp.) ß 1 =heat slope ß 2 =cold slope + ß 3 (pollution)+ ß 4 (influenza)+ ß 5 (day, PH) measured confounders + ß 6 (season)+ ß 7 (trend) unmeasured confounders
LAGS Heat impacts short: 0-2 days Cold impacts long: 0-21 days Vary by cause-of-death - CVD: prompt - respiratory: slow Should include terms for all relevant lags
LONDON, : LAGS FOR COLD-RELATED MORTALITY % INCREASE IN MORTALITY / ºC FALL IN TEMPERATURE DAYS OF LAG ALL CAUSE CARDIOVASCULAR RESPIRATORY NON-CARDIORESPIRATORY
SANTIAGO: COLD-RELATED MORTALITY CARDIO-VASCULAR DISEASE
SANTIAGO: COLD-RELATED MORTALITY RESPIRATORY DISEASE
SANTIAGO: COLD-RELATED MORTALITY ALL CAUSES
LAG: 0-1 DAYS HEAT LAG: 0-13 DAYS COLD Threshold for heat effect Threshold for cold effect
CONTROLLNG FOR SEASON TEMPERATURE MORTALITY SEASON Infectious disease Diet UNRECORDED FACTORS Human behaviours XX
Moving averages Fourier series (trigonometric terms) Smoothing splines Stratification by date Other… METHODS OF SEASONAL CONTROL
EFFECT OF INCREASING SEASONAL CONTROL Gradient of cold-related mortality, London
SEASONAL MORTALITY, GB
Month-to-month variation in mortality (adjusted for region and time-trend) accounted for 17% of annual all-cause mortality but only: - 7.8% after adjustment for temperature % after adjustment for influenza A counts - 5.2% after adjustment for both SEASONAL FLUCTUATION IN MORTALITY, GB
FUTURE IMPACTS
Seasonal mortality pattern, Delhi Daily deaths
Delhi, India: Average annual pattern of temperature, rainfall and daily mortality (data for all years, averaged, by day of year) 1st Jan1st July Jan 1 July 1 Dec Daily deaths Daily temperature Monthly rainfall Temperature Deaths McMichael et al, in press
Heat-related mortality, Delhi Relative mortality (% of daily average) Daily mean temperature /degrees Celsius Temperature distribution
Health impact model Generates comparative estimates of the regional impact of each climate scenario on specific health outcomes Conversion to GBD ‘currency’ to allow summation of the effects of different health impacts GHG emissions scenarios Defined by IPCC GCM model: Generates series of maps of predicted future distribution of climate variables RISK ASSESSMENT FOR CLIMATE CHANGE
EXTRAPOLATION (going beyond the data) VARIATION (..in weather-health relationship -- largely unquantified) ADAPTATION (we learn to live with a warmer world) MODIFICATION (more things will change than just the climate) BUT FOUR REASONS TO HESITATE…
Source: Checkley et al, Lancet 2000 Daily hospitalizations for diarrhoea Daily temperature HOSPITALIZATIONS FOR DIARRHOEA, LIMA PERU Shaded region corresponds to El Niño event
SOI (Southern Oscillation Index) Hales and Woodward, 1999 El Nino years La Nina years
Changes in population - Demographic structure (age) - Prevalence of weather-sensitive disease Environmental modifiers Adaptive responses - Physiological habituation (acclimatization) - Behavioural change - Structural adaptation - PH interventions CHANGING VULNERABILITY
Provide evidence on short-term associations of weather and health ‘Robust’ design Repeated finding of direct h + c effects Some uncertainties over PH significance Uncertainties in extrapolation to future (No historical analogue of climate change) SUMMARY: TIME-SERIES STUDIES
INTERMISSION…
TIME SERIES ANALYSIS FOR STUDIES OF WEATHER AND HEALTH Part 2
HARVESTING
FRAILTY MODEL General population Frail population, N t Death DtDt ItIt N t = N t-1 + I t - D t-1
IDEALIZED SCHEMA MORTALITY HEAT A B weaker absent strong correlation Period of averaging
ALL-CAUSE MORTALITY
ALL CAUSE MORTALITY: HEAT DEATHS Lag PERCENTAGE INCREASE IN MORTALITY PER ºC BELOW COLD THRESHOLD * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
CUMULATIVE EXCESS RISK OF HEAT DEATH AS A FUNCTION OF INCREASING LAG: LONDON
CUMULATIVE EXCESS RISK OF HEAT DEATH AS A FUNCTION OF INCREASING LAG: DELHI
CUMULATIVE EXCESS RISK OF HEAT DEATH AS A FUNCTION OF INCREASING LAG: SAO PAULO
CUMULATIVE EXCESS RISK OF HEAT DEATH: DELHI, SAO PAULO, LONDON
UNCERTAINTIES IN FUTURE HEALTH IMPACTS
(1) EXTRAPOLATION
MEAN DAILY TEMPERATURE / degrees Celsius MORTALITY (% of annual average) ? HEAT DEATHS Monterrey, Mexico
(2) VARIATION
ISOTHURM STUDY: annual pattern of temperature, rainfall and daily mortality (data for all years averaged, by day of the year) LJUBLJANABUCHARESTSOFIADELHI MONTEREYMEXICO CITYCHIANG MAIBANGKOK SALVADORSAO PAULOSANTIAGOCAPE TOWN
Mortality (% of annual average) Mean daily temperature in degrees Celsius Daily mortality in relation to mean temperature during preceding two days
Daily mortality in relation to mean temperature during preceding two weeks
(3) ADAPTATION
Threshold for heat impacts Maximum daily mean temperature Thresholds for heat-related mortality: 12 lower- & middle-income cities Positive slope suggests adaptation
Risk of death relative to annual minimum 10-week moving average Day of year 1Jan1Apr1Jul1Oct31Dec High standardized heating costs Low standardized heating costs Seasonal variation in deaths from cardiovascular disease by cost of home heating. England,
CARDIOVASCULAR MORTALITY IN RELATION TO HOME HEATING: ENGLAND, Standardized indoor temp. /deg Celsius Mortality (deaths/day) Winter:non-winter ratio* WinterNon- winter UnadjustedAdjusted for deprivation 1 < (1080) 0.6 (1568) 1.39 (1.28,1.50) 1.38 (1.16,1.63) (973) 0.6 (1580) 1.24 (1.15,1.35) 1.24 (1.05,1.47) (869) 0.5 (1442) 1.21 (1.11,1.31) 1.21 (1.02,1.44) (957) 0.6 (1569) 1.22 (1.13,1.32) 1.23 (1.04,1.46) (1055) 0.7 (1906) 1.11 (1.03,1.20) 1.13 (0.96,1.34) * All ratios adjusted for region
Slope becomes shallower if home is warmer Mortality Outdoor temperature / degrees Celsius
(4) EFFECT MODIFICATION
Table. Change in population health and deaths attributable to cold over the 20 th century Percentage of deaths by age & cause: Period years years 65+ years 38.5% 32.0% 29.4% 13.3% 40.5% 46.1% 4.9% 31.4% 63.7% 1.5% 18.8% 79.7% Cardiovascular Respiratory Other 12.1% 18.9% 69.0% 27.9% 20.0% 52.1% 33.3% 14.1% 52.6% 42.3% 14.0% 43.7% Percent of deaths attributable to cold 12.5 (10.1, 14.9) 11.2 (8.40, 14.0) 8.74 (5.93, 11.5) 5.42 (4.13, 6.69)
Temperature (degrees Celsius)
years65+ years Population attributable fraction (PAF) of deaths from cold PAF decade
IMPLICATIONS FOR MONITORING HEALTH IMPACTS OF CLIMATE CHANGE
METHODOLOGICAL ISSUES Gradual change Year to year fluctuation Secular trends Modifiers - physiological acclimatization - structural and behavioural adaptation - specific protection measures Attribution
1. Measurement of trend in disease rates 2.Measurement of trend in attributable disease: direct method 3.Application of climate- disease relationships to measured changes in climate: indirect method Confounded by secular trends: un-interpretable unless v. specific marker Based on analysis of (short-term) climate- disease relationships Depends on understanding effect modification or assumption of its absence MONITORING
Deaths in June & July, London, Deaths in June & July Year Days over 27ºC Deaths in June & July Days over 27 Celsius
Deaths attributable to heat, London, Percent attributable to heat year Days over 27ºC Heat deaths Days over 27 Celsius
Band of historical climatic variability Average Global Temperature ( O C) Year Low High Central estimate = 2.5 o C (+ increased variability) IPCC (2001) estimates a o C increase This presents a rate-of-change problem for many natural systems/processes
NEEDED EVIDENCE MITIGATION Evidence for change that benefits health & lowers emissions Reduction in greenhouse gas emissions/energy use Social, economic and technological changes ADAPTATION/PREPAREDNESS Evidence that can influence health in short and longer term Understanding of weather-health > climate-health relationships Vulnerability in terms of impacts, geographical distribution and population characteristics Public protection through: public health system (short-medium term) infrastructure, adaptation
CONTACT DERTAILS Sari Kovats Paul Wilkinson Public & Environmental Health Research Unit London School of Hygiene & Tropical Medicine Keppel Street London WC1E 7HT (UK) Tel: +44 (0) Fax: +44 (0)