Intended for healthcare professionals

Editorials Temperature is a complex confounder and may be inadequately accounted for

Air pollution and short term mortality

BMJ 2002; 324 doi: (Published 23 March 2002) Cite this as: BMJ 2002;324:691
  1. Enid Hennessy, lecturer in medical statistics (e.m.hennessy{at}
  1. Department of Environmental and Preventive Medicine, Wolfson Institute, Queen Mary's School of Medicine and Dentistry, London EC1M 6BQ

    The widely held belief that European levels of air pollution might seriously affect human health has been hard to verify. Most studies investigate the immediate effects on mortality, comparing day to day variations in atmospheric pollutants, temperature, and humidity with variations in deaths. High and low temperatures increase mortality while the effects of cold at least are prolonged1 and temperature is associated with pollution. Most investigators attempt to adjust for this confounding, although methods vary. Because both relations are complex, simple adjustments for confounding may be inadequate and some or all the apparent effects of pollutants may be indirect associations caused by the effects of temperature. A recent paper by Keatinge and Donaldson now questions whether this is true for some recent studies such as those from the “air pollution and health—a European approach” (APHEA) project, citing sulphur dioxide as an example.2

    Several reports suggest that atmospheric sulphur dioxide is associated with short term mortality, although the variability in results needs explanation. The APHEA project was designed to investigate this by, among other things, using several years of daily measurements from 15 European cities and standardised analyses.3 Katsouyanni et al concluded from it that rising sulphur dioxide levels increased mortality in western but not eastern Europe, although their results also varied.4 The protocol for temperature adjustment for the APHEA project was deliberately simple and used only one measurement—that day's temperature, or the temperature on one of the previous three days, or an average temperature.

    There is always a trade off in statistical analyses when dealing with potential confounding factors between using a simplification of a potential confounder and using a more complex method to model its effects. Those investigating air pollution are aware that temperature affects mortality, and many go as far as adjusting for its non-linear confounding effects in their analyses. However, even this may be inadequate if they do not adjust for the delayed effects of temperature more completely, because some pollutants like sulphur dioxide are associated with air temperature over time. For example, high levels of sulphur dioxide tend to occur at the end of longer than average cold spells. Such a non-random association could introduce bias into the results.

    Keatinge and Donaldson, in their study on air pollution and mortality in London, avoided the need to adjust for non-linear effects by restricting data to the daily mean temperature range 0°C to 15°C but did adjust for the delayed effect of temperature.2 They found little or no effect of sulphur dioxide, suggesting that the positive result from the APHEA project might be biased by inadequate adjustment for temperature. Using their data, they compared the two methods. Sulphur dioxide had little or no effect with theirs (which included immediate and delayed effects on mortality—up to 24 days), while it appeared significantly harmful when the APHEA protocol was used. This was true whether or not other meteorological terms were included. They cite Mackenbach's Amsterdam study, which showed similar reductions.5 It is not possible to rule out chance as an explanation for the change of effect in the two studies,2, 3, 4, 5 but if oversimplification of temperature causes substantial bias, some previous results could be wrong.

    Keatinge and Donaldson showed two other relevant results.2 Firstly, the relation between pollution and temperature depended on time, as the level of sulphur dioxide was generally higher at the ends of periods of cold calm weather. This should be expected because air pollution rises during periods of calm stable air, which encloses sources of pollution, and usually falls following a change in weather. Slow moving or static high pressure areas (anticyclones) with their temperature inversions typify these conditions and cause rising pollution if they enclose a source of pollution.6 They are also associated with colder winter and warmer summer weather. The days with high sulphur dioxide levels at the end of these periods of anticylonic weather will have raised mortality compared with those earlier because of the accumulation of more delayed effects of temperature. Thus if adjustment for temperature is oversimplified the extra deaths will be attributed to sulphur dioxide. Secondly, cold reduced deaths on the day (which Keatinge and Donaldson ascribe to cutaneous vasoconstriction reducing demands on failing hearts), although net mortality was increased over time. Within their temperature range of 0°C to 15°C the day after high pollution is likely to be milder and will seem to have an increased risk of death. This will probably be ascribed to the previous day's level of sulphur dioxide if adjustment is insufficient.

    To check whether simplifying the effects of temperature would cause substantial bias, I analysed simulated models of excess mortality, temperature, and sulphur dioxide.7 In the simulations temperature affected mortality for five days and sulphur dioxide had no effect at all. If the temperature on the day and the previous four were included in the analysis, sulphur dioxide was shown correctly to have no effect, but if the analysis was restricted, as in the APHEA studies, to measures from only one day each, then a range of effects for sulphur dioxide was shown. Many of these were relatively large and harmful. It is likely that some of the APHEA results for sulphur dioxide are biased. The APHEA2 project uses more terms to deal with delayed effects of temperature8 but similar analyses using these temperature terms still showed some biases, albeit smaller.

    These potential biases might partly explain some features of the APHEA results: the optimum delays found for sulphur dioxide and temperature by season and the relative size of winter and summer effects. As patterns of climate and pollution will affect biases, this could also explain some of the variability in results.

    The general message for time series analyses of short term mortality is that bias can occur if confounders have embedded effects in time which are inadequately adjusted for. Specifically for air pollution and short term mortality, it would be advisable to include both immediate and several delayed effects of temperature and analyse by season. The effects on mortality of delayed effects and non-linearity of temperature, and continuing effects over extended periods of extreme weather, probably differ by season.

    A reanalysis of the APHEA data or investigation of the association in time between pollutant and temperature should clarify the short term effects of pollutants. The subsequent questions remain difficult to answer. For example, do pollutants merely bring forward deaths (if any) by days or weeks or are some deaths advanced by years? Also unanswered is the question of whether air pollution causes cancer or other diseases in the long term.


    1. 1.
    2. 2.
    3. 3.
    4. 4.
    5. 5.
    6. 6.
    7. 7.
    8. 8.