Methods

The project involved the initial development of analytic approaches for first-stage and second-stage analyses of the time-series data and the subsequent application of the resulting methods to the time-series data. With regard to the first-stage analysis, the various investigative groups had used conceptually similar approaches to the key issues of controlling for temporal confounding and temperature; however, specific methods differed. Consequently, the investigators needed to establish a standard protocol, but one that would be linked to prior approaches. Based on exploratory analyses and simulation studies, a first-stage analysis protocol was developed that used generalized linear models (GLM) with either penalized splines (PS) or natural splines (NS) to adjust for seasonality, with 3, 8, or 12 degrees of freedom (df) per year and also the number of degrees of freedom chosen by minimizing the partial autocorrelation function (PACF) of the model's residuals. For hospitalization data, the approach for model specification followed that used for mortality, accounting for seasonal patterns, but also, for weekend and vacation effects, and for epidemics of respiratory disease. The data were also analyzed to detect potential thresholds in the concentration-response relationships. The second-stage analysis used pooling approaches and assessed potential effect modification by sociodemographic characteristics and indicators of the pollution mixture across study regions. Specific quality control exercises were also undertaken. Risks were estimated for two pollutants: particulate matter-10 pm in aerodynamic diameter (PM10) and ozone (O3).

Results

The first-stage analysis yielded estimates that were relatively robust to the underlying smoothing approach and to the number of degrees of freedom. The first-stage APHENA results generally replicated the previous independent analyses performed by the three groups of investigators. PM10 effects on mortality risk estimates from the APHEA2 and NMMAPS databases were quite close, while estimates from the Canadian studies were substantially higher. For hospitalization, results were more variable without discernable patterns of variation among the three data sets. PM10 effect-modification patterns, explored only for cities with daily pollution data (ie, 22 in Europe and 15 in the US), were not entirely consistent across centers. Thus, the levels of pollutants modified the effects differently in Europe than in the United States. Climatic variables were important only in Europe. In both Europe and the United States, a higher proportion of older persons in the study population was associated with increased PM10 risk estimates, as was a higher rate of unemployment-the sole indicator of socioeconomic status uniformly available across the data sets. APHENA study results on the effects of O3 on mortality were less comprehensive than for PM10 because the studies from the three regions varied in whether they analyzed data for the full year or only for the summer months. The effects tended to be larger for summer in Europe and the United States. In the United States they were lower when controlled for PM10. The estimated effect of O3 varied by degrees of freedom and across the three geographic regions. The effects of O3 on mortality were larger in Canada, and there was little consistent indication of effect modification in any location.

Conclusions

APHENA has shown that mortality findings obtained with the new standardized analysis were generally comparable to those obtained in the earlier studies, and that they were relatively robust to the data analysis method used. For PM10, the effect-modification patterns observed …