Data and data sources

Based on ICD code of diagnosis for inpatient admission, all cardiovascular (ICD10 codes I00–I52) and respiratory (ICD10 codes J00–J99) admissions between 1 January 1995 to 30 December 2004 (i.e. a total of 3,652 days) were obtained from the two public hospitals in Nicosia (i.e. Nicosia General and Makarios Hospital) with information on gender, age, date of admission and discharge as well as whether inpatient was resident in the district of Nicosia. The daily volume of all-cause admissions in the same period was obtained from the Cyprus Statistical Services, aggregated in 8 age- and sex-strata – males and females, aged 4 or less, 5–14, 15–64 and 65 or older. Hourly measurements of PM₁₀ and O₃ were available from two ambient air quality monitoring stations: (a) a local traffic-representative station located centrally at the Nicosia General Hospital and (b) a station reflecting background levels in the rural location of Ayia Marina, approximately 40 km from Nicosia. Continuous monitoring of coarse particles was performed with TEOM (Tapered Element Oscilling Micro-balance) instruments. Data from the rural station were available only from 1997 for ozone and 1999 for PM₁₀, reducing the study period to 8 and 6 years respectively. Thus, the main analyses presented here were based on a single exposure series as recorded at the traffic-representative station. Completeness of the main data series of air pollutants was generally satisfactory and concurred with previous practice, including the protocol used in APHEA. Commonly, missing values are replaced using a weighted average of values from other monitoring stations on that day. With only one other station, located rurally and thus not necessarily representative of exposure in the capital, imputation of missing values was not considered not least because availability of data did not cover the full period of investigation and would replace only a small fraction of the missing values. The background station was mainly used to aid in the identification of dust-storm days, since due to its rural location it is not affected by local traffic pollution. Nevertheless, the effect on the observed estimates of using the background station to define exposure was considered in sensitivity analyses. Daily averages of air and dew point temperature, relative humidity, wind speed, precipitation and barometric pressure based on hourly measurements of Thermohygrographs (i.e. instantaneous values) were obtained from the Cyprus Meteorological Services.

Identification of dust-storm days

Dust storm events affect Cyprus at least a few days a year resulting in extreme PM concentrations, which sometimes may persist for a few consecutive days. To construct a calendar of such events, an iterative approach was used whereby starting with a pool of days (N = 773) with at least 1 hourly measurement of PM₁₀ higher than 150 μg/m³ recorded at Nicosia Central or higher than 100 μg/m³ at the rural station (in both cases, 2 standard deviations (SD) away from the mean of hourly values on the logarithmic scale), paper-form meteorological records from the main International Airport at Larnaca (50 km from Nicosia) were reviewed to identify those days when a meteorological observer (as part of their hour-by-hour coding of weather and visibility conditions) noted poor visibility due to "dust in suspension" at any part of a given day, which did not refer to hazy conditions or dust from local sources i.e. a result of re-suspension. Candidate days were then cross-checked against a number of data-based criteria to assess the extent to which levels of PM₁₀ on those days where indeed uncharacteristic, including (i) daily average levels of PM₁₀ higher than 2 SDs away from the mean as recorded at the rural station, not as prone to traffic pollution and/or (ii) levels of PM₁₀ higher than 2 SDs away from the mean in the centre of Nicosia after excluding hours of pick traffic and/or (iii) days identified as outliers (2 SDs away from the predicted value) using a predictive model of levels of PM₁₀ based on levels the previous day and adjusting for weather variables including temperature, barometric pressure, precipitation and wind speed. The data-based criteria were not used to identify dust-storm days per se but to correct the original calendar which was based only on meteorological observations (i.e. either confirm, invalidate or propose additional days) with the main aim of categorizing days into confirmed events, where the criteria were in agreement with the meteorological observation and suspected events, where all criteria suggest an event that might have been overlooked or coded otherwise by the observer. Possible markers of dust storm events used in the past, such as low carbon monoxide levels [14], reduced visibility range [15], or based on investigation of aerosol optical depth [16, 17], were not electronically available. Finally, in accordance to previous practice [16, 18–20], backwards wind trajectories for up to 4 days ending in Cyprus (35°N, 33°E) on the day and about the time the meteorological observation was made were used to track the possible source of each identified event in the Sahara or Arabian peninsula using the National Oceanic and Atmospheric Administration (NOAA) HYSPLIT model.

Statistical analyses

Only days with at least 12 hourly measurements were used to calculate daily averages of air pollutants (PM₁₀ and O₃) as well as daily 8-hour maximum moving average for O₃. The effect of changes in levels of air pollutants on the number of daily admissions was investigated in Poisson regression models (a) as linear terms, expressed as percentage increase in mean number of daily admissions per 10 μg/m³ or 10 ppb increase in levels of PM₁₀ and O₃ respectively, (b) across quartiles of increasing levels of PM₁₀ (after including and excluding dust storm days) and O₃ to assess non-linearity of effects and (c) to include a categorical variable for dust storm days and restrict the estimation of a linear effect to non-storm days only (i.e. Dust Storm Day (= 0 or 1) + PM₁₀ daily average × [1-Dust Storm Day]). The magnitude of effect on these days was then compared to the effects seen across quartiles of all non-storm days with increasing levels of PM₁₀. In accordance to previous practice, to ensure that extreme values of PM₁₀ (in this case, most likely to be a result of air blown dust) would not influence the estimation of linear effects, the main analyses excluded days with average PM₁₀ concentrations greater than 150 μg/m³ (25 days). Furthermore, to assess the extent to which associations persist at lower ambient levels of air pollution, analyses were repeated restricted to days with daily averages of PM₁₀ less than 100 μg/m³ or 75 μg/m³ (the European standard at the time). Finally, to assess and correct for overdispersion, or extra-Poisson variation in the data, models were repeated using negative Binomial models (i.e. an overdispersed Poisson distribution). Model fit was assessed by inspection of the overdispersion parameter, the model deviance as well as patterns and magnitude of the residuals.

Generalized Additive Models (GAM) with natural splines were used to remove long-term seasonality (starting with the practical choice of 40 degrees of freedom to capture 4 seasons over the 10 year period) and penalized cubic splines to control for possible non-linear effects of the meteorological variables on the outcome (with a maximum of 5 degrees of freedom). In order to control for short-term patterns of admissions, day of the week was included in the models as a categorical variable. Before including the air pollution variables in the model, a base model was constructed to remove seasonality i.e. long-term trend and weekly patterns. Minimising the absolute value of the sum of the partial autocorrelation function was used to assess the appropriateness of the degree of smoothing. To avoid oversmoothing and, thus, eliminating patterns due to the exposure under study, windows below 2 months were generally not considered [21]. The weather variables as well as the appropriate lags for these to include in the models were then chosen by minimizing the Unbiased Risk Estimator (UBRE) and/or the Akaike's Information Criterion (AIC). All models were checked for remaining autocorrelation by examining plots of the partial autocorrelation function and, if necessary, sensitivity of the inferences was assessed in autoregressive Poisson models. The final model controlled for long- and short-term trend, temperature on the same day as well as the two previous days (lags 1 and 2) and relative humidity on the same day. Wind speed, precipitation and barometric pressure were not considered as confounders of the association between air pollutants and hospitalization. The effect of ozone was assessed before and after controlling for the effect of PM₁₀.

Same-day and lagged exposure (up to 2 previous days) were considered. Due to lack of statistical power, distributed lag models were not considered [22]. Similarly, due to the small number of daily cause-specific admissions, analyses were only stratified by gender (all ages combined) or age (i.e. younger/older than 15 years of age). Models were repeated to include interaction terms between levels of pollutants and season to investigate evidence for effect modification either (a) during cold and warm months indexed by monthly average temperature levels or (b) on days when average daily temperature was higher or lower than 20 or 30°. Finally, where possible (i.e. cause-specific investigation), all analyses were repeated to exclude non-Nicosia residents since these may represent transfers from other hospitals and, as such, can dilute any effect since day of admission may not accurately reflect day of exposure. Data manipulation was performed in STATA SE 9.0. Analyses were performed using the MGCV package in the R software (R 2.2.0).

Main findings

Short-term changes in PM₁₀ increase the risk of same-day hospitalization for all-cause and cardiovascular causes in Nicosia hospitals. An effect on respiratory health was observed only during the warm season. In contrast to the effects seen with same-day levels of PM₁₀, only lagged effects were observed with levels of ozone. These associations appeared independent of the effect of levels of PM₁₀ and, perhaps, stronger for cardiovascular causes. More interestingly, there was also evidence to suggest that risk of hospitalization was higher on dust storm days. While a statistically significant association was observed only for overall admissions, in all cases, effects were at least comparable in magnitude to those seen on non-storm days with the highest levels of PM₁₀, particularly for cardiovascular causes.

Limitations

Several studies have shown that the effect of air pollution on hospitalization is stronger for certain cardio-respiratory conditions [24] and among certain sub-groups of the population, such as children, the elderly or people with a recent history [25–28]. The small number of daily events in a city the size of Nicosia provides limited statistical power to permit a finer age-, sex, or cause-specific analysis. This is, in fact, portrayed in the often low precision of the estimates and, thus, wide confidence intervals. Furthermore, multiple testing across sub-groups or different lags might produce some spurious associations. For this reason, the sub-group analyses were a priori restricted to gender (only)- and age (only)-specific comparisons and lagged exposure for up to 2 days. Due to lack of statistical power in a small sample, distributed lag models were not considered. Models investigating effect modification and non-linear effects across quartiles are only complimentary to the main analysis, nevertheless necessary, in the first case to highlight important differences by season and, in the case of the latter, to provide a basis of comparison (and a similar unit of measurement) between dust storm and non-storm days. The observed patterns appear not only internally consistent but, as a result of the long period of investigation (i.e. 10-years), the magnitude of estimates in a small city the size of Nicosia seem to be in agreement with those observed in large studies elsewhere.

As with all hospital data, there might be to some extent misclassification of the cause of admission, particularly in people with both respiratory and cardiovascular pathologies. However, it is unlikely that such misclassification is temporally related to levels of air pollutants and, thus, can only bias our estimates towards the null. With only one air quality station in Nicosia in operation for the full 10-year period of investigation, the analyses presented here are based on a single exposure series. Thus, replacing missing values in about 10% of days (and an additional 2–3% days purposefully excluded due to availability of less than 12 hourly measurements) was not possible. Nevertheless, the data series is generally longer than in similar time-series studies. Furthermore, similar patterns were observed when analyses were repeated for the reduced period for which data were available from the background station. Finally, identification of dust-storm days was mainly based on visibility observations and coding practices of a meteorological observer. Subjective in nature as it may be, a series of data-based criteria were used to validate the observations. It is likely that some dust storms, especially those of milder intensity, might have been missed out. However, it is unlikely that days with high levels of PM solely due to traffic sources were considered as dust-storm days. At any rate, the observed effect persisted when restricting inferences to the group of days for which meteorological observations and measurements at the three sites were in agreement.

The effect of short-term changes in air pollutants on hospitalisation

Even though at limited power due to the small number of daily events in a city the size of Nicosia, the long period of investigation has allowed for some positive (and statistically significant) associations between levels of particulate matter and risk of hospitalization to be observed. Estimates seem consistent with the size of effects seen across several European cities, particularly for cardiovascular causes (i.e. 0.4%–1.4% per 10 μg/m³ increase in PM) [29, 30]. Surprisingly, other than the warm season, no overall effect was observed with respiratory admissions. While several studies have shown associations between levels of PM and respiratory mortality [6, 10], there have been some inconsistent results in the case of hospital admissions [31], with some studies only showing associations in certain sub-groups [32] or for certain respiratory causes, such as COPD and asthma [33]. Furthermore, admissions during the colder months are mainly driven by viral respiratory infections. Even though models adjusted for the observed seasonal pattern with higher respiratory admissions in the colder season, it is possible that not explicitly controlling for these causes (due to lack of data) has masked a possible effect in the colder months, particularly in children (aged less than 15). It is also important to note that, in the case of respiratory admissions, a striking weakly pattern was observed with a large dip in admissions on Tuesdays (2 daily admissions on average) on either side of high volume on Mondays and Wednesdays (9 daily admissions). With the exception of the elderly (aged 65+), similar patterns were observed in all age and sex groups. While it was not clear whether this was due to reduced bed availability, there was some indication that people admitted on Monday were more likely to be discharged before Wednesday (42.5%) while those admitted on Wednesday were more likely to be kept until the following Monday (as many as 37% and only 13% for any longer). The extent to which this atypical weekly pattern has produced a discrepancy between actual need and admission and, thus, contributed to the lack of any strong association with overall respiratory admissions is not known.

No positive effects were observed with levels of PM₁₀ the 2 previous days. While lagged or cumulative effects have commonly been observed between levels of PM and mortality in several studies [34], it is not uncommon for the strongest associations to be observed with same-day levels in the case of hospital admissions [12, 35]. Perhaps, even more so in the case of Cypriot cities where, with no referral system, direct access to specialist health care may be considerably easier than in larger European or US cities. A number of person-based studies have now used average levels of air pollutants the 24 hours preceding an event (rather than same calendar day which would include hours after as well as before the event) and, at least amongst susceptible populations, have shown effects in as little as 6 hours prior to an arrhythmia and even 2 hours prior to a myocardial infarction [36–39].

In contrast to the same-day association with levels of PM, only effects of lagged exposure to ozone were observed. Negative associations with same-day levels of ozone have previously been observed in several of the 23 APHEA cities, an effect which persisted even in the summer period in at least 4 of the cities, namely Rome, Paris, Tel Aviv and Valencia [40]. While explanations for this are not clear, it is thought to be a product of the relationship between ozone (a secondary pollutant produced by photochemical reactions) and traffic-related primary pollutants such as nitrogen oxides (NO) and volatile organic compounds (VOC) which can reduce ozone concentrations at least at the local scale. Thus, high ozone concentrations in stations located in central parts of cities, such as in this case, may reflect low levels of local traffic pollutants or good dispersion conditions, which can lead to negative associations with health indicators, at least in the short-run, whereas it is common for stronger associations with ozone to be subsequently observed at lagged intervals. A similar pattern of negative same-day associations, albeit much weaker, was observed with levels of ozone as recorded at the rural station, not influenced by local pollution. Finally, while, there have generally been some inconsistent findings for the association between cardiovascular morbidity and levels of ozone, with some reporting positive [41], no or even negative associations [42], effects of a similar magnitude have been previously reported between levels of ozone two days before admission for cardiovascular causes in London [43]. In general, neither the fact that stronger associations with lagged rather than concurrent exposure nor that stronger effects for cardiovascular than respiratory admissions were observed seem inconsistent with previous findings [43–45].

Effect modification by season

Seasonal differential effects of changes in levels of air pollution on the risk of hospitalization have been previously reported, particularly with respects to stronger effects of PM on respiratory health during the warm season such as those observed here [32, 43, 46]. Other studies, however, have found little evidence of differential effects by season [47, 48]. It is uncertain whether such differential effects carry biological plausibility, i.e. a synergistic effect of air pollution and temperature amplifying people's response to lower levels of air pollutants than normally, or it simply, reflects an increased proportion of time spent outdoors, and thus higher exposure, on warm days [49]. In APHEA, for instance, the stronger effects of PM₁₀ on total mortality observed in the warmer than colder cities persisted when latitude was used instead of actual temperature, thus, it was proposed that ambient measurements in warmer places may represent the average population exposure more accurately than in colder places where people do not spend as much time outdoors [21]. Here, other than some weak evidence of strongest effects on cardio-respiratory morbidity on the warmest days (>30°), generally, more pronounced differential effects were observed across cold-warm months rather than across cold-warm days (based on daily mean temperature).

In the light of evidence of ozone effects on mortality restricted to the warm season across 23 APHEA cities, the stronger association between cardiovascular admissions and ozone during the colder season observed here appears surprising [40]. However, some similar patterns have been observed elsewhere, particularly amongst the elderly [50]; possible explanations for this pattern are unclear and it might simply be a chance finding. Nonetheless, in a recent cross-sectional study in Korea, it was reported that measures of blood pressure, a risk factor for cardiovascular disease, were significantly associated with PM₁₀ levels only in the warm season while the reverse was true for O₃, with associations only during the cold season [51].

The effect of dust storms

Long-range transport of Saharan dust across the Mediterranean into southern Europe [52, 53], and less frequently as far north as the British Isles [54], is well established. Consistent with a recent examination of these events [23], our analysis also suggests that there has been a rise in their frequency, at least half of which seem to occur during the spring months and can last for as many as four consecutive days. Unlike the mineral and chemical composition as well as transport patterns of Saharan dust, the possible health effects of these natural events have not been extensively studied. There is some modest evidence from Taiwan and Korea of adverse effects of wind-blown dust from the Mongolia/China desert on both cardiovascular and respiratory health, some times lasting for up 3 days after the event [55–57]; reported associations were, however, not always statistically significant.

In contrast, a study of 17 episodes with a high concentration of coarse (crustal-derived) but not fine particles during a six-year period in Spokane, Washington, revealed no evidence of an increased risk of death on dust days [14]. Similarly, an investigation into an unusual event of transported dust over the Atlantic to Greater Vancouver, Canada, has shown no effect on hospital admissions [58]. Some studies have, in fact, reported strengthening of the observed health effects when windblown dust days were excluded from the analyses [59] or reduced health effects on windy days [60], in both cases suggestive that crustal-derived particles are more benign that those from anthropogenic processes. In our analyses, we have not observed an attenuation of effects by including dust-storm days. On the contrary, we found evidence of increased admissions on dust-storm days of similar magnitude to the effects seen on non-storm days with the highest concentrations. Of course, the possibility of air-borne dust containing particles of anthropogenic sources can not be excluded since mega-cities, such as Cairo, are commonly on the path of these events. Alternatively, some support for the adverse effects of particles due to natural sources observed here comes from a recent study of the biological content in dust transferred to Haifa, Israel during similar events in 2004–2005 that revealed both an increase in the concentration of airborne microorganisms and a change in the usual content of fungal population in PM samples, both which are thought to affect human health [19]. Similar findings have been reported from aeromicrobiological analyses of samples on the Turkish coastal town of Erdemli [61].