Statistical methods
The time-series analyses were conducted by using gener-
alized additive models with nonparametric smoothers to
control for seasonal patterns. A smoother is a tool for sum-
marizing the trend in one variable, in this case the number
of visits to the general practitioner, as a function of another
variable, in this instance time. This method allows for very
flexible control of unusual seasonal patterns such as those
observed in the allergic rhinitis series. The span of the
smoother was varied to control the amount of smoothing
carried out on the time series. The aim was to select the span
that removed long-term seasonality but left short-term pat-
terns, since they may be caused by fluctuations in air pollu-
tion levels. The amount of smoothing required varied
between series; therefore, to determine the amount of
smoothing needed, a relatively large span was used initially
and the model diagnostics examined. Successive reductions
in the smoothing window were then made, and individual
smoothers were created for more problematic periods of the
series. Model diagnostics were reassessed at each step.
Goodness of fit of the statistical model was assessed from
the model residuals, the magnitude of the dispersion param-
eter, the partial autocorrelation function, and the model
deviance.
Dummy variables were used to allow for day-of-the-week
effects. Temperature and humidity were included in the
model after various diagnostic plots of the seasonally
adjusted model residuals were considered against different
lags, both single and cumulative, of the meteorologic vari-
ables. Depending on which were more appropriate, either
parametric functions or broad smoothers of the meteorologic
variables were used to model the associations. Variation in
the practice population over the 3-year period, counts for
lagged allergic pollen measures, and the daily number of
consultations for influenza were all adjusted for in the “core”
model if necessary. Once all of these potential confounding
variables had been included, the air pollution indicator was
added to complete the statistical model. After we allowed for
overdispersion and autocorrelation if necessary (13), Poisson
generalized additive models (quasi-likelihood) regression (or
robust regression if deemed more appropriate) was used to
determine the relative risk of consultation associated with an
increase in the pollution measure. All statistical analyses
were carried out by using S-PLUS software (14).