Short-Term Mortality Rates during a Decade of Improved Air Quality in Erfurt, Germany

Background Numerous studies have shown associations between ambient air pollution and daily mortality. Objectives Our goal was to investigate the association of ambient air pollution and daily mortality in Erfurt, Germany, over a 10.5-year period after the German unification, when air quality improved. Methods We obtained daily mortality counts and data on mass concentrations of particulate matter (PM) < 10 μm in aerodynamic diameter (PM10), gaseous pollutants, and meteorology in Erfurt between October 1991 and March 2002. We obtained ultrafine particle number concentrations (UFP) and mass concentrations of PM < 2.5 μm in aerodynamic diameter (PM2.5) from September 1995 to March 2002. We analyzed the data using semiparametric Poisson regression models adjusting for trend, seasonality, influenza epidemics, day of the week, and meteorology. We evaluated cumulative associations between air pollution and mortality using polynomial distributed lag (PDL) models and multiday moving averages of air pollutants. We evaluated changes in the associations over time in time-varying coefficient models. Results Air pollution concentrations decreased over the study period. Cumulative exposure to UFP was associated with increased mortality. An interquartile range (IQR) increase in the 15-day cumulative mean UFP of 7,649 cm−3 was associated with a relative risk (RR) of 1.060 [95% confidence interval (CI), 1.008–1.114] for PDL models and an RR/IQR of 1.055 (95% CI, 1.011–1.101) for moving averages. RRs decreased from the mid-1990s to the late 1990s. Conclusion Results indicate an elevated mortality risk from short-term exposure to UFP. They further suggest that RRs for short-term associations of air pollution decreased as pollution control measures were implemented in Eastern Germany.

Research Ambient concentrations of particulate matter (PM) have been consistently associated with daily mortality (Analitis et al. 2006;Dominici et al. 2005;Health Effects Institute 2003;Pope and Dockery 2006;Zanobetti et al. 2003). Associations between ambient concen trations of nitrogen dioxide or carbon mon oxide and daily mortality have been observed (Samoli et al. , 2007, but the causality of the NO 2 effects is being debated [World Health Organization (WHO) Europe 2006].
In recent years, air pollution concentra tions have been reduced by emission controls and fuel replacement. Accountability of these measures is of major concern for the regulat ing agencies as well as for the regulated enti ties. This is of particular interest because some areas in the United States and in Europe are still out of compliance with enacted stan dards (U.S. Environmental Protection Agency 2004) or suggested guideline values (WHO Europe 2006).
A small number of studies have been con ducted that fit within this research framework. Examples are studies on the banning of coal sales in Dublin, Ireland (Clancy et al. 2002), the reduction of sulfur in fuels in Hong Kong, China (Hedley et al. 2002), and traffic restric tions during the 1996 Olympic Games in Atlanta, Georgia, USA (Friedman et al. 2001) and the 2002 Asian Games in Busan, Korea . A recent analysis of data from the Harvard Six Cities study investigated whether a decline in mortality rates is largest in cities with the largest reduction in longterm average PM (Laden et al. 2006). Lately, a study investigated the association between particulate air pollution and mortality in the United States during a period when several key particulate related air pollution control programs were implemented (Dominici et al. 2007).
Political changes in Central and Eastern Europe have resulted in the restructuring of the Eastern bloc industries, improved emission controls, and a changed car fleet (Acker et al. 1998). The improved emission control led to a complete fuel replacement and an exchange of brown coal for natural gas in power plants and in domestic heating. All those changes resulted in improved air quality in this region within a decade (Ebelt et al. 2001) and have provided an opportunity for a natural experiment to evaluate the health impacts of air pollution. Heinrich et al. (2002) assessed the impact of declines of total suspended particulates and sul fur dioxide in Eastern Germany in the 1990s on the prevalence of nonallergic respiratory disorders in children.
We designed the present study to inves tigate the associations between selected cri teria pollutants [NO 2 , CO, and PM < 10 µm (PM 10 ) and PM < 2.5 µm (PM 2.5 ) in aerodynamic diameter], ultrafine particle number concentrations (UFP), and daily mortality over a 10.5year period after the German unification, with a particular empha sis on changes in relative risks (RRs) for daily mortality in association with these pollutants, as they changed during the study period.

Material and Methods
Study area and period. We conducted the study in Erfurt, the capital of the state of Thuringia, Germany, from 1 October 1991 to 31 March 2002. The city of Erfurt has a pop ulation size of approximately 200,000 inhab itants and is surrounded by mountain ridges of 100-150 m on three sides and highrise buildings on the fourth side. Because of an administrative reform in 1994, a number of autonomous communities were incorporated into the city area of Erfurt [see Supplemental Material, Figure 1 (http://www.ehponline. org/members/2008/11711/suppl.pdf)]. The city area including those communities meas ures 21 km from north to south and 22.4 km from east to west. However, about 90% of the inhabitants of Erfurt live within a rect angular area of 5 km × 3 km around the old city center.
Mortality data. We obtained copies of death certificates without the name and address of the decedents from local health authorities to comply with the rules of the German data privacy law. We excluded deaths of infants (< 1 year of age) and nonnatural deaths [International Classification of Diseases, 9th Revision (ICD9;WHO 1975) codes ≥ 800 and 10th Revision ICD10 (WHO 1993) codes ≥ S00]. Data collection methods and quality control mechanisms are described elsewhere (Wichmann et al. 2000).
For the analysis of air pollutants, for which data were available from 1991 onward, we considered deaths occurring within the old city limits of Erfurt [see Supplemental Material, Figure 1 (http://www.ehponline.org/mem bers/2008/11711/suppl.pdf)]. For the analysis of air pollutants in the period 1995-2002, we also included deaths in the incorporated communities. This strategy provides estimates comparable with earlier analyses of subsets of the data set presented here (Stölzel et al. 2003(Stölzel et al. , 2007Wichmann et al. 2000).
Air pollution and meteorologic data. We obtained daily mean concentrations of NO 2 and CO from a staterun network monitoring station for the entire study period. During the winter of 1991-1992 and from September 1995 onward, we sampled the particle size distribution at a research monitoring site located around 1 km south of the city center [see Supplemental Material, Figure 1 (http:// www.ehponline.org/members/2008/11711/ suppl.pdf)]. The measurement station can be classified as an urban background site and had a distance of 40 m from the nearest major road. We measured sizespecific particle num ber concentrations by an aerosol spectrom eter as described elsewhere (Pitz et al. 2001;Tuch et al. 2003;Wichmann et al. 2000). For the present analysis, we computed daily means of UFP (size range, 0.01-0.1 µm) from the spectra. We obtained data for the num ber concentrations (NC) for three specific size ranges-0.01-0.03 µm, 0.03-0.05 µm, and 0.05-0.1 µm-for September 1995 to August 2001. We computed daily means of PM 2.5 assuming spherical particles of a mean density of 1.53 g/cm 3 (Pitz et al. 2001;Wichmann et al. 2000). Additionally, we collected PM 10 on a Harvard Impactor (Air Diagnostics and Engineering Inc., Harrison, ME, USA).
We imputed missing values in the UFP, PM 2.5 , and PM 10 time series using concurrent measurements. A detailed description of the imputation process can be found elsewhere (Peters et al. in press;Stölzel et al. 2007). Between 1 April 1994 and 1 February 1995, the NO 2 concentrations were unusually low and exhibited very little variation. Therefore, we excluded this period from the analyses.
We obtained daily mean air tempera ture and relative humidity from a site of the German Meteorologic Service located at Erfurt Airport 5 km west of the measurement station.
We calculated exposure lags up to 14 days for the air pollution data. In addition, we cal culated the means of lags 0-5 and 0-14 for the air pollution data and the means of lags 0-1, 0-2, and 0-5 for the meteorologic variables, if at least half of the relevant lags were available.
Other data. We obtained data on influ enza epidemics from the Arbeitsgemeinschaft Influenza [AGI (German Influenza Working Group) 2003] in the form of a weekly doctor's practice index for each winter season (October through April). This index indicates the rela tive deviation of the number of doctor visits because of acute respiratory symptoms com pared with a background level averaged for the whole of Germany.
Statistical analysis. Statistical model. We analyzed data using generalized semiparamet ric Poisson regression models. We used natu ral cubic and penalized regression splines to allow for nonlinear confounding effects. We considered constant as well as timevarying asso ciations between pollutants and daily mortality. We built confounder models separately for the two analysis periods, 1991-2002 (gaseous pollutants and PM 10 ) and 1995-2002 (UFP andPM 2.5 ), without including any air pol lutants. As potential confounders, we consid ered a global trend over calendar time, seasonal and weekday variations, influenza epidemics, and air temperature and relative humidity. We selected models by minimizing Akaike's Information Criterion (Akaike 1973) and the absolute value of the sum of the partial auto correlation function . To ensure sufficient adjustment for season and meteorology, we forced longterm time trend and sameday air temperature into all mod els. We considered lags 0-2, the mean of lags 0-1, and the mean of lags 0-2 for the weather variables; for the doctor's practice indexes, we assessed shifts of up to ± 3 weeks, because the influenza epidemics may have reached their peaks in Erfurt at another time than in the whole of Germany.
In the final confounder models [see Supplemental Material, Table 1 (http://www. ehponline.org/members/2008/11711/suppl. pdf)], we readjusted the number of degrees of freedom (df) for the smooth function of time trend, because many of the meteorologic variables exhibit seasonal patterns themselves and hence capture part of the observed seasonal trends in the outcome (Touloumi et al. 2004).
In the last step of the analysis, we added air pollutants separately to the models and estimated associations linearly. To investigate cumulative associations between air pollutants and daily mortality counts up to 14 days after exposure, we used polynomial distributed lag (PDL) models (Schwartz 2000;Zanobetti et al. 2003). We constrained the lag coefficients to follow a thirddegree polynomial of the lag number. We obtained cumulative estimates as the sum of the estimated coefficients for any given lags in the PDL models. Additionally, we investigated the associations between 6day or 15day averages of the pollutants and mortality.
For estimating pollution-mortality asso ciations over different periods, we replaced the assumption of a timeconstant overall pollution effect by using an interaction term approach from which the air pollutant has been removed as a main effect. This means that we included indicator variables for the periods and multi plicative linear terms of the pollutant and the period indicators. This approach can be seen as a simple timevarying coefficient model that results in a step function or as a nested approach to investigating effect modification by period (Van Ness and Allore 2006). We used a likelihood ratio test to determine whether there were indeed differences between periods.
Alternatively, we adapted statistical tools developed by Brezger and Lang (2006) for Bayesian varying coefficient models. Specifically, we estimated timevarying associations of the pollutants by modeling the effect estimate as a smooth function of time trend β poll = f poll (t).
The smooth effect f poll (t) was modeled using a Bayesian adaptation of penalized Bsplines (Brezger and Lang 2006). A more detailed description of this modeling approach can be found elsewhere (Peters et al. in press); for a similar approach, see Lee and Shaddick (2007).
We analyzed data using the package "mgcv" in the statistical software R (R Development Core Team 2003) and using BayesX (Brezger et al. 2005). We present effect estimates as RRs for mortality together with 95% confidence/ credible intervals (CIs) based on an increase in air pollution concentrations from the first to the third quartile [interquartile range (IQR)].
Sensitivity analyses. To explore the robust ness of the models, we performed sensitivity analyses using different values of smoothness for the functions of time trend and air tempera ture. Furthermore, we used a categorical vari able for all days of the week instead of a Sunday indicator. For the analysis period 1995-2002, we performed a sensitivity analysis using the same confounder model as for period [1991][1992][1993][1994][1995][1996][1997][1998][1999][2000][2001][2002]. Moreover, we investigated the associa tion between daily mortality and UFP adjusting for other pollutants in twopollutant models. We finally investigated the exposure-response relationship between air pollutants and mortal ity. We replaced the linear term of the pollut ant concentrations with a fixed 3df regression spline. We used a likelihood ratio test with 2 df that compares the original main model with the smoothed model, and visual inspection to assess whether the smoothed exposure-response curve resembles a straight line. Additionally, we com pared different values of smoothness by the gen eralized crossvalidation score as provided in R.

Results
Mortality data. We collected 17,713 death certificates that met the inclusion crite ria. On average, we observed 4.6 deaths per day. Around 10% of all cases occurred in the outlying communities that were incor porated in 1994 [see Supplemental Material, Figure 1 (http://www.ehponline.org/ members/2008/11711/suppl.pdf)].
Air pollutants and meteorologic data. There were significant changes in air pollution con centrations during the 10.5 years of observation. We therefore divided the full study period into three smaller periods: period 1, 1 October 1991 to 31 August 1995; period 2, 1 September 1995 to 28 February 1998; and period 3, 1 March 1998 to 31 March 2002 (Figure 1). NO 2 , CO, PM 2.5 , and PM 10 levels decreased continuously between period 1 and period 3 ( Figure 1, Table 1). However, PM 2.5 and UFP were measured only during the first winter of period 1. During period 2, the UFP remained stable and decreased only after 1999-that is, in the middle of period 3. All pollutants exhibited a pro nounced seasonal pattern, with higher con centrations during the winters and lower concentrations during summers (Figure 1). UFP was only moderately correlated with PM 10 and PM 2.5 (Spearman rank correla tion = 0.57 and 0.48), whereas PM 10 and PM 2.5 were highly correlated [Spearman rank correlation = 0.85; see also Supplemental Material,  Figure 2 and Table 4 present cumula tive RRs estimated for different periods. The associations between air pollutant concen trations and mortality were strongest for the years 1995 to 1998 and decreased afterward (Figure 2). Although gaseous pollutant con centrations generally did not exhibit strong associations with mortality (Tables 2 and 3), they also showed adverse effects during 1995-1998 (Table 4, Figure 2). Using the alternative approach with a smooth function to model timevarying associations produced estimates that are comparable in shape [see Supplemental Material, Figure 4 (http://www.ehponline.org/ members/2008/11711/suppl.pdf)].
Sensitivity analyses. We performed a num ber of sensitivity analyses (Table 5). Further reducing the df for the spline to model trend and seasonality did not change or only slightly decreased the risk estimates for UFP. Using a Table 2. Cumulative RRs of mortality per IQR increase of air pollutants in Erfurt, Germany, estimated with PDL models of lags up to 5 days with a third-degree polynomial, or estimated with means of lags 0-5.

Discussion
Economic and political changes and the adop tion of new technologies in Eastern Germany have resulted in clear improvements in ambi ent air quality. We observed the largest RRs for UFP. We further observed that RRs varied over time for some of the pollutants, which could not be explained by nonlinearity in the exposure-response functions. We found no significant associations for PM 2.5 and PM 10 . In general, associations between mortal ity and air pollution were lower at the end of the study period than during the 1990s. We observed the strongest associations in the transition period 1995-1998, when changes in source characteristics took place and the benefits of ambient air quality were not yet completely achieved. Previous studies have indicated an asso ciation between UFP and daily mortality, but evidence is limited. Forastiere et al. (2005) showed a significant increase of outofhospital coronary deaths in Rome, Italy, in association with an increase in sameday UFP. Previous analyses of data from Erfurt, Germany, for the periods 1995-1998 and 1995-2001 indicated delayed associations between UFP and daily mortality (Stölzel et al. 2007;Wichmann et al. 2000). We observed the highest associa tion between UFP and daily mortality with a lag of 4 days. This analysis showed that the RR estimates for the entire period were some what smaller than in the previous analyses for data of the years 1995-1998, which is consistent with the results of the timevarying analyses. The analyses presented in this study showed largest RRs between 1995 and 1998, which, by chance, was about the time period of the initial mortality study in Erfurt (Stölzel et al. 2003;Wichmann et al. 2000). UFP and a UFP subclass showed evidence for lin ear exposure-response relationships when we applied smoothing techniques, which adds to the consistency of the associations. Two pollutant models suggested that the observed association of UFP with mortality was not confounded by other pollutants.
The association between PM and allcause mortality has been consistently observed (Analitis et al., 2006;Dominici et al. 2005;Health Effects Institute 2003). Large multi center studies in Europe and the United States have reported effect estimates of 0.2-0.6% per 10µg/m 3 increase of PM 10 (Health Effects Institute 2003). Previous analyses of data from Erfurt pointed to an association of mor tality with fine PM mass (Wichmann et al. 2000). A 19.9µg/m 3 increase in the PM 2.5 concentration was significantly associated with 3.0% more daily deaths (same day). In the present study, we observed no significant associations for PM 10 or PM 2.5 . However, timevarying models indicated positive effect estimates for the period 1995-1998, whereas in the other periods the effect estimates were indistinguishable from the null.
Several other studies investigated daily timeseries data for time scales of exposure substantially longer than just a few days by using frequency domain loglinear regression (Dominici et al. 2003) or by using distributed lag models (Goodman et al. 2004; Zanobetti Table 4. Cumulative RRs (95% CIs) of mortality per period in association with air pollution in Erfurt, Germany, estimated with multiday moving averages.

Pollutant
Overall   . 2003). Similar to these studies, we found larger RRs of mortality associated with particulate air pollution for longer time scales of exposure (15 days) than at time scales of a few days (6 days). During times of drastic changes in con centrations of the exposures, in principle, two competing factors may explain a change in effect estimates. The first possibility is a non linear exposure-response relationship that converts to changes in the effect estimates over time as concentrations in the exposure change. The second possibility is a linear exposure-response relationship, but that the measured pollutants serve as indicators for changes in source composition or air pollu tion mixtures, so the associations with the outcome vary. We observed linear exposureresponse functions and variations in RRs over time using timevarying coefficient models. However, we cannot fully exclude the possi bility that we observed associations by chance due to the low statistical power.
Other studies have assessed whether changes in the levels or composition of the aerosol are associated with changes in health impacts (Clancy et al. 2002;Hedley et al. 2002). Laden et al. (2006) observed that mor tality rates declined largest in cities with the largest reduction in longterm average PM using followup data from the Harvard Six Cities study. An analysis using data of the U.S. National Morbidity, Mortality, and Air Pollution Study showed only a weak indica tion that the association between PM 10 and mortality declined during 1987-2000 and that this decline occurred mostly in the east ern United States (Dominici et al. 2007). A crosssectional study conducted in three areas of Eastern Germany showed a decreas ing prevalence of respiratory symptoms in schoolchildren along with declines of total suspended particulates and SO 2 in the 1990s (Heinrich et al. 2002). Results from the present study suggest that the association between air pollution and mortality declined during 1995-2002. We observed the largest declines for the trafficrelated air pollutants NO 2 , CO, and UFP, but we also observed a borderline significant decline for PM 2.5 . The biggest changes in mortality from period to period occurred when the local power plant was changed over from coal combustion to natural gas. Consequently, one may speculate whether the reduced chronic exposure to the coal combustion effluents may have contrib uted to the decline in daily mortality in the following few years.

Strengths and limitations.
Erfurt is a small city with < 5 deaths per day on aver age, which limits the statistical power of the analyses. However, because of the uniqueness of the long record of particle size distribution measures, it nevertheless serves as a natural experiment in which the impact of mobile source emissions, and of the transition from coalbased energy to modern technologies in energy generation, on ambient air pollution concentrations was recorded over a decade.
The use of a single monitoring site for the whole city of Erfurt may pose a limita tion to this study. However, this site has been demonstrated to be representative for the air quality within the city of Erfurt with respect to PM 10 and sulfate (Cyrys et al. 1998). One reason for the strong spatial correlation of the air pollutants in Erfurt is the geographic situation. Erfurt is confined with ridges on three sides and highrise buildings on the fourth side. As a result, days with reduced air exchange rates between the city area of Erfurt and the surrounding rural area occur more frequently than in other German cities. Days with increased levels of ambient air pollut ants are therefore more frequent than in other German cities.
Because UFP is mostly produced by local traffic, a greater spatial heterogeneity could be expected. However, concurrent measurements of UFP at different sites within one city often have shown good correlations despite differing magnitudes and suggest that a background site might well represent the exposure of the average population with respect to UFP if the site is carefully chosen (Aalto et al. 2005;Buzorius et al. 1999;Cyrys et al. 2008;Peters et al. 2005). Nevertheless, one would expect greater exposure misclassification by a single monitoring site for locally produced particles such as UFP than for regionally transported particles. Therefore, the fact that the present study did identify larger associations between locally generated particles and mortality than between PM 2.5 or PM 10 and mortality may point to an important role of locally produced particles on daily mortality. Furthermore, the area of Erfurt, including the surrounding communities incorporated in 1994, is small. The doubling of the city region in the course of the administrative reform in 1994 may have led to less precise exposure assessment for the inhabitants of the newly incorporated communities, as suggested by the slightly larger risk estimates obtained for the old city center. However, about 90% of the inhabitants of Erfurt live within a rectangular area of 5 km × 3 km around the old city center.
We used a variety of pollutants for the analyses, because different pollutants may point toward differing properties of the aerosol and also represent different sources of air pol lution. By testing a set of air pollutants, how ever, the possibility that some effects might have occurred by chance cannot be excluded. Because air pollutants are closely correlated, we considered especially consistent patterns in the data as actual effects. Moreover, we thor oughly adjusted for meteorologic confounder variables to rule out the possibility that the detected associations resulted from meteo rologic influences or seasonal differences. In the present study, we smoothed confound ing variables using penalized splines, instead of loess smoothers as applied by Wichmann et al. (2000), because of recent concerns about biased estimates of the standard errors due to concurvity when using loess smoothers in the statistical software package Splus (Ramsay et al. 2003). The results are internally consis tent and qualitatively agree with those from previous analyses over a shorter time period (Stölzel et al. 2003;Wichmann et al. 2000) with respect to UFP, despite the slightly dif ferent modeling strategy described above. The confidence limits for the RR estimates were somewhat smaller because of the larger amount of data. Additional sensitivity analyses indicated that our final model seemed to be conservative and stable with respect to the choice of the model parameters.

Conclusions
Results indicate an elevated mortality risk from shortterm exposure to UFP highlight ing the potential importance of locally pro duced particles. The study further suggest that RRs for shortterm associations of air pollution decreased as pollution concentra tions decreased and control measures were implemented in Eastern Germany. The expo sure-response functions were linear, and the concentration changes did not explain the variation in the coefficients.