Deliverable D_O-INT_3.12

Analysis of relative influence of individual health stressing factors

Josef Keder

October 2011

Introduction. 3

Study period and data used. 3

Long-term and seasonal component removal 5

Regression analysis. 7

Regression analysis results. 7

Conclusions. 11

References. 12

Introduction

The association of health impacts represented by mortality (e.g. daily number of deaths) or morbidity (connected with the daily number of hospital admissions) and health stressing factors like pollutants concentrations and meteorological factors such as very low or very high temperature or sultriness has been shown in continuously increasing amount of studies. The health impact indicators show very distinct long-term patterns and seasonality reflecting the seasonal behaviour of weather phenomena. Respiratory disease epidemics and other cold weather related events may lead to the increased winter mortality and, in contrast, summer temperature extremes are related to sharp, short-term increases in mortality during of the warm season.

Correlation on the long-time scales (seasons) confounds effects of the primary causes of health problem outbreaks (short-term air quality and/or stressing weather factors changes). Assessment of the relative influence of particular stressing factors on the health outcome is complicated by the seasonality as well. In this study, based on the statistical analysis of the time series of the health and weather data, the weather and air pollution factors reported as the mostly stressing where chosen. Using the non-parametric scatterplot smoothing methods the long-term components were removed from the time series and the effects were estimated by the Poisson regression models. Incorporation of the lagged data into regression models enabled estimates of the possible delay in particular stressing factors effects and to indicate the possible presence of the harvesting (mortality displacement) effect.

Study period and data used

Among the weather variables, effect of air temperature, humidity and wind speed on the human well-being and health is widely reported as the mostly progressive in the literature. However, treating them as separate variables in multiple regression analyses with a health outcome as the dependent variable ignores the natural association between weather elements. Their combined effect may have the most important influence on human health. A series of composite weather indices have therefore been developed using these meteorological variables as an input (e.g. Leung et al., 2008). An approach based on the composite index has been applied in this study.

For the purposes of this study, meteorological data on ambient air temperature, relative humidity and wind speed from man-serviced and automated station network operated by the Czech Hydrometeorological Institute (CHMI), were utilized. This network covers the whole territory of the Czech Republic. With the aim to exclude the influence of extreme conditions, only stations situated at height below 700 m a.s.l. were incorporated. Detailed information on the Czech station network, including maps, might be found on the CHMI website http://www.chmi.cz.

Health data, comprising daily hospital admissions and daily death counts for 2003 – 2008 period, split by patient’s domicile (region) and selected diagnoses (cardiovascular, cerebrovascular and respiratory diseases, according to International Statistical Classification of Diseases and Related Health Problems), were kindly provided by the Institute of Health Informatics and Statistics of the Czech Republic.

Study area and its administrative partitioning (regions and districts) are depicted on the map in Fig.1.

Fig. 1: Czech Republic (the study area) and its administrative partitioning

The web sites of the Czech Statistical Office, http://www.czso.cz/eng/redakce.nsf/i/home, provide comprehensive geographical end demographical information.

Example of health data tables (daily mortality data) provided by Institute of Health Informatics and Statistics, split according to ICD-10 (International Classification of Deceases), is shown on the Fig.2.

Fig.2: An example of health data tables provided by IHIS.

The composite effect of air temperature, humidity and wind speed has been involved using so called apparent temperature (AT) according Steadman (1994). According to Steadman’s formulas, mean daily values of apparent temperature were calculated for each day of the study period and each meteorological station used. Wind speed has been fixed to 3m/s, a long term mean for the Czech Republic. From the station data the territorially averaged values were determined.

As for air pollution data, suspended particle concentration are widely used in studies of air quality and health relationships. The mostly harmful health effects are assigned to them in Czech Republic as well. Daily means of PM10 suspended particle fraction measured at Czech air quality monitoring stations were used in this study framework. The stations description may be found on the CHMI website

http://portal.chmi.cz/files/portal/docs/uoco/web_generator/locality/pollution_locality/index_GB.html

Only the data from station located below 700 m a.s.l. were selected and used for the regional means calculations.

Long-term and seasonal component removal

For the long-term and seasonal component removal from the time series an approach recommended by Schwartz (2000) and Michelozzi et.al. (1998) has been utilized. The LOESS scatterplot smoothing method described by Schwartz (2000) has been selected as the most appropriate for this purpose. The span 120 days recommended in Schwartz’s paper has been selected for the long-term component filtering. An example of the long-term component of all-cause (total mortality) filtered by LOESS is presented on the Fig. 3.

Fig. 3: Daily count of the all-cause (total) mortality in Czech Republic (dots) with the filtered long-term and seasonal component (red line)

The deseasonalized short-term mortality data were obtained by subtracting of this long-term component form the mortality counts data series. These data already do not show any seasonal patterns.

Fig. 4: Daily count of the all-cause (total) mortality deseasonalized by the long-term component subtraction

Long term components and deseasonalized data for the apparent temperature and mortality counts split by particular diagnoses were calculated by the same way.

Fig. 5: Long-term components of daily count of the all-cause (total) mortality and apparent temperature

On the fig. 5 the courses of the long-term components of daily count of the all-cause (total) mortality and apparent temperature are depicted. The negative correlation of mortality peaks and apparent temperature troughs in winter time is clearly apparent. This correlation might represent a confounding factor in regression analysis when deseasonisation procedure would not be  applied.

Regression analysis

With the aim to derive quantitative relations between meteorological variables, PM10 concentrations and mortality rates a regression model has been constructed under following assumptions:

-        The deseasosanlized mortality counts (split to total, cardiovascular, cerebrovascular and respiratory death causes) were used as a health outcome (dependent variable).

-        The combined effect of air temperature, humidity and wind speed has been involved using apparent temperature, averaged over the CZ territory and deseasosanlized.

-        Daily PM10 concentrations averaged over the CZ territory were used as a variable representing air pollution.

-        AT and PM10 variables lagged (shifted) by 1 day from the lag 0 till 10 were used for the delayed effects and possible mortality displacement estimates.

With the assumption that mortality counts data follow the Poisson distribution, regression models of the form

ln(E(Mort)) = A + B*AT + C*PM10

were fitted using Poisson regression with the logarithmic linking function.

Regression analysis results

Regression coefficient recalculated in the form

100*(exp(B)-1)

100*(exp(10*C)-1)

may be interpreted as percentage changes of mortality associated with the increase of AT by 1 degree and PM10 concentration increase by 10 Âµg.m^-3. Usually these numbers are interpreted as relative risk. Relative risks for particular diagnoses are summarized in the following tables.

Table 1: Relative risks in % associated with the 1 degree increase of AT and 10 Âµg.m^-3 increase of PM10 concentration. Statistically significant values are marked bold.

Relative risks associated with the AT and PM10 increase for the particulate mortality causes related to lags 0 till 10 days are shown on the following graphs. Confidence intervals 95% are depicted as well.

Fig.6: Relative risks associated with the AT by 1 deg and PM10 by 10 µg.m^-3 increases for the total mortality related to lags 0 till 10 days. Confidence intervals 95% are depicted.

Fig.7: Relative risks associated with the AT by 1 deg and PM10 by 10 µg.m^-3 increases for the cardiovascular mortality related to lags 0 till 10 days. Confidence intervals 95% are depicted.

Fig.8: Relative risks associated with the AT by 1 deg and PM10 by 10 µg.m^-3 increases for the cerebrovascular mortality related to lags 0 till 10 days. Confidence intervals 95% are depicted.

Fig.9: Relative risks associated with the AT by 1 deg and PM10 by 10 µg.m^-3 increases for the respiratory mortality related to lags 0 till 10 days. Confidence intervals 95% are depicted.

Conclusions

From the Table 1 and figures 6 – 9 the following conclusion can be formulated:

-          The apparent temperature (AT) which aggregates the influence of air temperature, humidity and wind velocity as main meteorological stressing factors shows no negligible effect on mean mortality increase expressed as a relative risk. The AT effect is mostly pronounced for the total and respiratory mortality.

-          The highest effect of the AT occurred at the lag 0, with the exception of the cardiovascular mortality, where a slight increase at lag 1 may be seen.

-          Air pollution is substantially associated with mortality and represents a significant risk. This is mostly pronounced for total mortality and for respiratory mortality as well. The maximum relative risk value has been evaluated for the cardiovascular mortality at lag 0, but the confidence interval contains the zero value and the regression coefficient is not statistically significant. Nevertheless, high relative risk values for cardiovascular mortality were estimated for the lags 3 and 7.

-          The highest effect of the PM10 occurred at lag 0, with the exception of the respiratory mortality for which the lag 1 showed the highest effect.

-          Interesting patterns can be observed in changes of relative risk associated with PM10 depending on the lag size. After the initial decrease with the increasing lag value, the secondary maximum occurred for lag the lag 8 for total mortality, lags 8 and 9 for the cerebrovascular mortality and for the lag 6 in case of respiratory mortality. Such patterns are typical for the presence of the mortality displacement. Secondary maximum presence at lags 6 – 9 indicates that after subsiding of the initial mortality increase associated to harvesting the effect of concentration increase on mortality might be delayed by about 1 week.

References

Huynen, M.M.T.E,Martens, P., Schram, D., Weijenberg, M.P., Kunst, A.E.: The impact of heat waves and cold spells on mortality rates in Dutch population. Environ. Health Persp., 109, pp. 463 – 470, 2001

Leung, Y.K., Yip, K.M., Yeung, K.H.: Relationship between thermal index and mortality in Hong Kong. Meteorological Application, 15, pp. 399-409, 2008

Michelozzi, P., Forastiere, F., Fusco, D., Perucci, C. A., Ostro, B., Ancona, C., Pallotti, G.: Air pollution and daily mortality in Rome, Italy, Occup. Environ. Med., 55, pp. 605–610, 1998

Schwartz, J.: Harvesting and Long Term Exposure Effects in the Relation between Air Pollution and Mortality. American Journal of Epidemiology. 151(5):pp. 440-448, 2000.

Steadman, R.G.: Norms of apparent temperature in Australia, Aust. Met. Mag., 43, pp., 1-16, 1994