The prevalence rates of all asthma and asthma without hayfever are shown in Table . With an overall prevalence of 18%, boys had higher prevalence for asthma (21%) than girls (15%), consistent with other Canadian studies on the life course of asthma [66
]. The difference between the two sexes was more apparent in the younger age group, with boys having higher rates of asthma without hayfever than girls. Prevalence rates for asthma ever were similar to the larger study from which our population was drawn, but our sub-population had slightly lower rates of wheezing ever for the younger children (data not shown).
Prevalence of asthma ever, non-atopy related asthma ever without hayfever ever, and hayfever ever
In total, sixteen exposure surfaces were created. These included the surfaces derived using the Theissen polygon method for O3, NOx, and SO2, and a splined surface developed for PM10. Figures , and show the surfaces for PM10Spline, O3Theissen and NO2LUR. All surfaces, except O3, showed a pollution gradient within the city that followed the expected trend of higher intensities in the northeast near the industrial core, and decreasing pollution levels towards the outskirts of the city. The O3 surface followed the opposite pattern, with lower levels in the downtown area and higher levels at the edges of the city.
The different pollution metrics reflect both different sources and different approaches to modelling exposure. The SO2
Theissen polygon surface indicates the presence of point-source industries in the northeast end of downtown Hamilton. The Theissen polygon surfaces created discrete categories of pollution levels equal to the measurements obtained at the fixed MOE sites. As these are not smooth continuous surfaces, they may not accurately reflect the real variation of pollution – an inherent unavoidable characteristic of creating such polygons around the monitoring locations [67
]. The O3
Theissen surface had a similar categorizing effect. The PM10
Spline, on the other hand, may have over-smoothed the true variation of pollution, again, due to the nature of this interpolation technique. Both the kriged and LUR NO2
surfaces were based on a denser network of monitors within the city. With the large variation in concentrations measured by the monitors, the kriging methodology was not able to capture the full spatial variation without incorporating some unavoidable errors included in the estimation. The highest errors, however, tended to be outside the area encompassing the children's residence locations. Visually, the NO2
LUR surface appeared most heterogeneous, with the highest variation occurring around roads and densely populated areas of the lower city.
The concentrations of estimated pollutants were then assigned to the postal code of each child's residential address. Pollution exposures in each group were quite similar for most pollutants (Table ) with the exception of smaller ranges of PM10
exposure for girls and NO2
LUR exposures for boys. A correlation matrix was constructed for the independent variables (see Additional file 1
). The pollutants had low correlations, except for O3
, which were inversely related due likely to the scavenging effect of ozone by local sources of NO [68
]. As expected, the two measures of NO2
were correlated. The DI had a weak positive correlation with the pollutants, except with O3
where there was a weak negative correlation. Dwelling value and average income were highly correlated (r = 0.73) and followed very similar patterns in their correlations with pollutants. The rate of repair and percent old houses were also highly correlated (r = 0.72). To avoid introducing multicollinearity, only one of the two variables in each correlated set was retained for the multivariate analysis. DI and smoking did not have strong associations with the other variables, and thus were kept for further testing in the multiple regression models.
Average and range of pollution exposures
Bivariate logistic regression revealed positive, but insignificant, associations between pollution exposures and asthma outcomes when the whole population was tested. A detailed table is available in the online appendix (see Additional file 2
). The odds ratio (OR) for asthma with NO2
LUR, for example, was 1.02 per ppb (95% CI, 1.00–1.04) but the association was insignificant. Samples were stratified based on literature suggesting that differences exist between the sexes [69
], that rates differ by age, and that asthma and asthma without hayfever have different risk factors [31
]. When testing the predictive potential of atopy related status for asthma, a positive association was found; namely that children with hayfever symptoms were more likely to have asthma than those without hayfever symptoms (OR = 3.03; 95%CI, 2.20–4.17). After testing interactions between the pollutants, atopy and subgroups, we found effects suggestive of an interaction between hayfever and pollutants in all girls for NO2
LUR (p = 0.156). The power to test for interactions in epidemiological studies is often poor, resulting in researchers missing important interactions due to lower power [70
]. As noted by Selvin [71
], relaxing the type 1 error p value from the traditional 5% to 20% is a common approach in epidemiological studies, one that can allow for interaction tests in studies that are not powered for effect modification. In this instance, we had substantive reasons to test for interaction, and the sub-group analysis indicates that girls are more susceptible than boys. Given this, the literature of subgroup interaction and the empirical evidence in our data, we subsequently stratified the sample into subgroups by age, sex, and by age and sex. It is important to note that all subgroups were investigated; however, due to the large number of tested associations between the subgroups, and the voluminous resulting tables, only the significant results for the susceptible groups are included in this paper to avoid detracting from the main findings by listing tables of insignificant effects.
Tables , , and show the associations for the stratified analysis conducted for the non-atopy related asthma population within the subgroups. Asthma without hayfever was associated with NO2LUR for all girls and older girls. We also ran trivariate logistic regressions on the significant associations identified in the bivariate tests for asthma without hayfever (see Table ). The effects of pollutants remained robust. NO2LUR retained significance with asthma without hayfever in all girls for each confounding variable. For the subpopulation of older girls, the odds ratios were generally stronger after adjustment for potential compositional and contextual variables. We used these regressions to test for the effect of the compositional and contextual variables on the percent change in the regression coefficients of the models. Testing the most robust group of associations (NO2LUR and all girls) identified the deprivation index (DI) and rate of repair as the variables that reduced the regression coefficient by more than 10%. These were retained as the confounding variables for the multiple variable logistic regressions.
Odds ratios of bivariate associations between asthma without hayfever and both NO2LUR and confounding variables within subgroups of all, younger and older children+
Odds ratios of bivariate associations between asthma without hayfever and both NO2LUR and confounding variables within subgroups of all girls and boys+
Odds ratios of bivariate associations between asthma without hayfever and both NO2LUR and confounding variables within subgroups of younger girls and boys+
Odds ratios of bivariate associations between asthma without hayfever and both NO2LUR and confounding variables within subgroups of older girls and boys+
Odds ratios of trivariate regressions between asthma without hayfever, NO2LUR and confounding variables within subgroups with significant bivariate associations+
We also tested the effect of co-pollutants on our models (see Table ). For the populations of all girls and older girls, the effect of NO2LUR was larger after adjusting for SO2, PM10 and O3. Calculated for a 1-unit increase in NO2, the odds ratio for asthma without hayfever among all girls was 1.46 times (after controlling for PM10, SO2, O3, DI and rate of repair), and 2.71 times greater among older girls.
Co-pollutant models for asthma without hayfever, controlling for DI and rate of repair+
Sensitivity analyses were conducted with a generalized linear model (GLM) with a natural spline smoother [65
]. The coefficient changed slightly from 0.128 to 0.130 when the smoother was applied to 10 degrees of freedom (df), and to 0.129 for 20 df (p < 0.05). Using the natural spline smoother, increasing spans indicated a more localized analysis. This sensitivity analysis lends further support to the notion that confounding probably does not bias the coefficients as the effects were robust.
For further sensitivity analysis, we tested the alternative indicators of atopic conditions (eczema and runny nose not associated with a cold) to assess whether the selection of indicator made a difference in the air pollution and asthma relationship. The results were sensitive to the selection of other indicators of atopic conditions, as they were positive for the most part, but were no longer significant.
We also assessed the sensitivity of the association of wheeze ever and current wheeze with the same symptom indicators of atopic conditions, and the pattern of effects was similar to that observed for asthma. After controlling for confounders and copollutants, NO2LUR remained significant with wheeze ever without hayfever (OR = 1.13, 95%CI, 1.01–1.23) and current wheeze without hayfever (OR = 1.28, 95%CI, 1.06–1.55) for all girls (OR = 1.15, 95%CI, 1.00–1.31) and older girls (OR = 1.35, 95%CI, 1.10–1.66).