Comparison of time series and case-crossover analyses of air pollution and hospital admission data.

Abstract

Time series analysis is the most commonly used technique for assessing the association between counts of health events over time and exposure to ambient air pollution. Recently, case-crossover analysis has been proposed as an alternative analytical approach. While each technique has its own advantages and disadvantages, there remains considerable uncertainty as to which statistical methodology is preferable for evaluating data of this type. The objective of this paper is to evaluate the performance of different variations of these two procedures using computer simulation. Hospital admission data were generated under realistic models with known parameters permitting estimates based on time series and case-crossover analyses to be compared with these known values. While accurate estimates can be achieved with both methods, both methods require some decisions to be made by the researcher during the course of the analysis. With time series analysis, it is necessary to choose the time span in the LOESS smoothing process, or degrees of freedom when using natural cubic splines. For case-crossover studies using either uni- or bi-directional control selection strategies, the choice of time intervals needs to be made. We prefer the times series approach because the best estimates of risk that can be obtained with time series analysis are more precise than the best estimates based on case-crossover analysis.