Title: Bayesian Metaanalysis and Health Impact Evaluation in the Study of Heat Effect on Mortality
1Bayesian Meta-analysis andHealth Impact
Evaluation in the Study of Heat Effect on
Mortality
Mini-symposium Environmental Epidemiology
Thursday 2 August 2007
- Annibale Biggeri, Michela Baccini
- Department of Statistics, University of Florence,
- Biostatistics Unit, CSPO, Florence, Italy
2motivating example the PHEWE study
Epidemiological daily time series
Poisson regression
- PHEWE study (Assessment and Prevention of acute
Health Effects of Weather conditions in Europe)
enrolled 15 cities, about 30 million people on
calendar years 1990-2001. - Generalized estimating equation approach was used
in season-specific analysis, with AR-1 error
structure and parametric spline or segmented
regression to model apparent temperature effects.
- Distributed lag and time-varying coefficient
models were fitted. - Effect Heterogeneity was expected, a priori two
regions were considered (Mediterrean vs Others).
3PHEWE study 1990-2001City-specific curves (GEE
natural cubic spline)Maximum Apparent
Temperature lag03 - All natural deaths - Summer
4PHEWE 1990-2001 threshold-slope model
in red Mediterranean cities
5comments
- A pattern was evident, but with large variability
among cities - (we do not discuss here city-specific modelling
issues - restricting the analysis by season and using
Generalized Estimating Equation approach with
AR(k) working correlation lead to - assure appropriate error structure for
consistency of GEE estimates in case of few long
time series (Xie and Yang, 2003) on the basis of
a formalized exploratory analysis (dynamic
regression models Pankratz, 1991 Chiogna and
Gaetan, 2005) - avoid use of sandwich estimator of standard
errors (Guo et al. 2005) - exposure effects modelling by Cubic regression
spline with 1 knot every 8 C and straight line
above a city-specific threshold)
6outline
- The PHEWE study documented
- heterogenous effect of heat on
- mortality across Europe
- Hierarchical Bayesian approach is used
- to estimate random-effects meta-analytic
summaries - to perform Multivariate random effect
meta-analysis and - to obtain shrunk estimates of city-specific
parameters - Montecarlo approach is used in Impact assessment
- sampling from city-specific posterior
distributions and - modelling climate change effects via
probabilistic distributions
7Statistical modelling - 1
- In PHEWE, with exception of the description of
the exposure-response function (where we used a
fixed-effect approach), the city-specific results
were combined using - Hierarchical Bayesian random-effects
meta-analysis. - Separated meta-analyses performed for two a
priori defined geographical regions - Mediterranean cities
- Athens, Barcelona, Ljubljana, Milan, Rome, Turin,
Valencia. - North-Continental cities
- Budapest, Dublin, Helsinki, London, Paris, Praha,
Stockholm, - Zurich.
8Hierarchical Bayesian random-effects meta-analysis
- First level for each cities (c1,,C) we have a
vector of coefficients Y and - the associated
variance-covariance matrix V (known)
- Second level the vector of true coefficients
follows a multivariate Gaussian - distribution with vector
of means ß and var-cov matrix A
-
-
9- Inference is based on full posterior
distributions - we use
- in the multivariate case direct sampling from
the marginal posterior distribution of ß (Everson
and Morris, 2000 TLNise library of R) - In the univariate case MCMC methods, using
WinBugs 1.4
marginal posterior global effect
marginal posterior het variance
marginal posterior city-specific effect
10rationale
- Hierarchical Bayesian meta-analysis was adopted
here because - Heterogeneity was a study objective (ML
underestimates it) - Multivariate effects are easy to model
- City-specific posteriors or shrinkage
distributions which updated first stage estimates
can be obtained - Uncertainty can be incorporated in Impact
evaluation phase
11PHEWE study 1990-2001Meta-analytic curves
Maximum Apparent Temperature lag 03 - All
natural deaths - Summer
12ER percent variation in mortality for 1C
increase in apparent temperature.
threshold
slope
13- Random effect meta-analysis is difficult to
interpret because - it assumes that the effect is not unique but
varies among study - the overall effect is a weighted average of
study-specific effects, the weights include
heterogeneity among studies - a paradox arises because the greater the
heterogeneity the more uniform the weights - The purpose of a meta-analysis changes in
presence of heterogeneity - the interest lies in explaining it
- Ecological bias is an issue
- the meta-analysis becomes an observational study
- (the value of proof is weakened)
14effects
PHEWE 1990-2001 posteriors threshold-slope
heterogeneity among cities
15- Large heterogeneity for threshold among
Mediterranean cities was a major result (see
posterior for het-variance) - (varying acclimatization ?
- east-west gradient ?)
16Multivariate meta-analysis Time-varying
coefficients model
- The temperature effect is allowed to vary
smoothly over time within year (Peng et al., 2004)
17sensitivity of the slope to season
definitionmeta-analytic curves (time-varying
coefficient models)
some evidence of greater effect in early summer
18Multivariate meta-analysis Distributed lag models
-
- Where the coefficients are constrained by a
polynomial function of grade 5, with lags
spanning 0,40 (Almon, 1965). -
19lagged effect / harvesting meta-analytic curves
20Statistical modelling 2Impact estimates
Attributable Deaths - YoLL
- To assess the impact of high summer temperatures
on the health of European urban populations we
estimate Attributable Deaths and Years of Life
Lost associated with high summer temperatures,
through secondary analysis of mortality versus
apparent temperature functions derived as part of
the PHEWE study. - For each city we used a probabilistic MonteCarlo
approach - observed temperature by day distributions
- risk distributions (posteriors)
21model specifications and assumptions heat
attributable deaths
- Let us restrict now the attention to attributable
number of deaths (AD). - We assumed that all population was exposed to the
same daily maximum apparent temperature. - We assumed that variations in maximum apparent
temperature under the threshold do not affect
mortality and that the effect is linear above the
threshold. - The same temperature threshold by city (t0c) is
used for all age classes. Thresholds are based on
posterior city-specific estimates. - Above-threshold slopes (bc) are obtained from
age-class specific meta-analytic posterior
distributions. - y is the death counts
22observed temperature by day distributions PHEWE
1990-2001
23Posterior distributions of city-specific
thresholdsMaximum Apparent Temperature lag03
SummerPHEWE study 1990-2001
24Posterior distributions of city-specific slopes
by age groupsMaximum Apparent Temperature lag03
SummerPHEWE study 1990-2001
25computing heat attributable deaths
- We use a Monte Carlo approach for each city, we
sampled values from the city-specific posterior
distributions of the slope and from the
city-specific posterior distributions of the
threshold obtained from the Bayesian
meta-analysis. - Then, for each sample, we calculated a time
series of daily number of attributable deaths
from the observed time series of daily maximum
apparent temperature at lag 0-3 and daily number
of deaths, according to the following formula - zero otherwise
26- Independence between threshold and slope was
checked by sensitivity analysis. - We obtained 10000 samples of AD time series by MC
simulation. - Separated AD evaluations were produced by city
and age group (15-64, 65-74, 75) and scenarios
(a total of 100001536). - For each city, total number of AD was produced by
summing AD over the three age classes.
27Heat alternative scenarios
- Temperature
- pseudo-series for
- Hottest year
- Coolest year
- Second-to-coolest
- Second-to-hottest
- Average increase
- in Temperature 1C
28details on building scenarios
- We defined scenarios selecting
- the summer with the highest mean level of
apparent temperature. - the summer with the lowest mean level of apparent
temperature. - the second to hottest day
- the second to coldest day
- for example, for the third scenario we
considered the hypothetical summer constituted by
the second to hottest 1st April, the second to
hottest 2nd April.the second to hottest 30th
Sept. - Exposure scenarios were also defined adding to
each observed daily apparent temperature a random
term from a Normal distribution
29PHEWE study 1990-2001 Mean number of observed
attributable deaths by calendar day, by city.
30PHEWE study 1990-2001Heat attributable deaths /
year, people 75yrs oldobserved temperature
distribution
31PHEWE study 1990-2001Heat attributable deaths
low - high temperature scenarios
80 Credibility bands
32PHEWE study Attributable deaths observed
apparent temperature 1990-2001 versus predicted
series by 1C mean increase scenario
AD attributable deaths AveT average
temperature d days above threshold
33summary
- The PHEWE study documented
- heterogenous effect of heat on
- mortality across Europe
- Hierarchical Bayesian approach is used
- to estimate random-effects meta-analytic
summaries - to perform multivariate random effect
meta-analysis and - to obtain shrunk estimates of city-specific
parameters - Montecarlo approach is used in Impact assessment
- sampling from city-specific posterior
distributions and - modelling climate change effects via
probabilistic distributions
34acknowledgments
- The PHEWE study group (EU DGR V Framework)
- Michelozzi, Kirchmayer, Katsouyanni, Biggeri et
al. Assessment and prevention of acute health
effects of weather conditions in Europe, the
PHEWE project background, objectives, design. - Environmental Health, 2007 6,12
- Baccini, Biggeri, Accetta, Katsouyanni et al.
Effects of Apparent Temperature on Summer
Mortality in 15 European Cities Results of the
PHEWE Project. (submitted) - Baccini, Kosatsky, Biggeri. Estimating deaths
attributable to summer heat through climate and
response scenarios anchored on an 11-year, 15
European city, temperature/mortality function
(unpublished manuscript)