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

Abstract

‘There’s so much pollution in the air now that if it weren’t for our lungs there’d be no place to put it all’, thus reads a famous Robert Orben quote. Indeed, even though ambient pollution levels experienced now in Western nations are much reduced compared with the devastating smog episode times of the 1940s and 50s, in many large cities of low-income countries the urban air pollution situation has worsened due to population growth, industrialization, and increased vehicle use. Further to this, much of the epidemiological evidence conducted in the West points to an adverse pollution effect even at levels well below current government standards. Traditionally, many of these epidemiology studies have investigated the short-term effects of air pollution on health using time series data, and, in recent years, the favoured statistical approach has been to employ generalized additive models (GAM) for this purpose. Over the past 12 months, however, it has emerged that GAM programming in statistical packages such as the widelyused S-Plus software may result in unstable estimates due to inadequate convergence 1 and may underestimate standard errors due to concurvity in the data. 2 Consequently, some doubts have been cast over the validity of previous findings, and a clear need has developed to investigate the robustness of such results. In a very interesting and timely paper, Fung and colleagues present an analysis of air pollution and hospital admission data using the GAM approach with stricter convergence criteria, and then compare the results with those obtained from alternative approaches such as the use of parametric natural cubic splines as part of a generalized linear model, and the application of the case-crossover design. 3 In general, the authors demonstrate consistency in their results using the different methods. Other studies are now beginning to find that effect estimates may reduce as a result of more appropriate convergence, but in general the qualitative conclusions remain the same. 4 The S-Plus default convergence parameters have already been revised in the new S-Plus version, and revisions of the GAM software, allowing ‘exact’ calculations of the standard errors, have now been developed. 5 This ‘GAM crisis’ has highlighted the need and indeed demonstrated the intention of the air pollution community to continue refining their methods. As these computational issues become resolved, other methodological concerns have come to the fore. It has been suggested that inadequate control of weather may explain some of the pollution effect, 6 and, perhaps most importantly, no firm consensus has been reached on the most appropriate degree of control for time. In the absence of such agreement, it is perhaps best to follow suit with Fung et al. and consider the sensitivity of the results to a variety of different specifications.