Strengths and limitations of the study
The main strengths of the present study are the large and almost complete number of cases (N = 4 904), the complete study base, and the high-quality exposure assessment of residential outdoor NOx.
Even so, this study had limitations. As in all epidemiologic studies of this kind, and although our exposure assessment was of high quality, there was exposure measurement error leading to exposure misclassification.
Zeger and colleagues discuss three components of exposure measurement error in time-series on air pollution data:.[
30] 1) Error due to aggregation of exposure, 2) Difference between average personal exposure and true ambient level, and 3) Difference between true and measured ambient level. Since we study long-term rather than short-term exposure, we reformulated "3", to be the difference between modelled and true ambient NO
x-concentrations and added another component of the error, 4) Error caused by migration.
1) Aggregated (500 × 500 spatial resolution) rather than individual exposure estimates leads to a Berkson type error that should not lead to bias in linear models and generally only marginal bias in log-linear models.[
30]
2) Concentrations outside a residence (where we model NO
x) do not necessarily correlate well with levels inside, or with actual exposure.[
31] We did not have data on time spent at home or on other sources of exposure, except for data on smoking status for the second-phase subjects. Although smoking is a major source of indoor air pollution, adjusting for it did not alter the effect estimates substantially (Table
3). Other differences between between actual and ambient exposure most likely cause bias towards the null, but it is hard to estimate the magnitude. By adjusting adequately for socio-economic factors, which has a complex relationship to air pollution,[
18] and likely a large influence on other potential exposures to air pollution, bias caused by this component of measurement error might be reduced.
3) The validation between modelled and measured levels indicated a measurement error of classical type, typically also yielding bias towards the null. If we assume that the true effect is OR = 1.10 for a 10 μg/m
3 increase in NO
x, and assume that the measurements accurately reflect the true levels, we can apply the formula by Armstrong[
32,
33] to calculate the expected OR as
= 1.09, where the ratio in the exponent is given by the measured ("true") variance, V
T = 7.7
2 and the modelled (observed) variance, V
obs = 8.3
2. Thus, for modest effects, the bias caused by this error component does not seem to be substantial
4) The proportion that had not migrated (estimated to 78%) within 10 years before the stroke year was similar between cases and second-phase controls, again implying bias towards the null.
If we conservatively assume that the error in the 22% that migrated is completely of classical type and the variance in those 22% is double to the true variance (
V
T
) in the 78% that did not migrate. For a true effect of OR = 1.10 per 10 μg/m
3 increase in NO
x, we get the expression:
, a moderate bias.
In summary, the error components for which we could estimate the impact did not seem to substantially influence the effect estimates. However, we cannot rule out that differences between actual personal exposure and ambient exposure may have yielded more substantial bias. Also, exposure measurement error would yield more substantial bias if the true OR was larger than 1,1, say for example 1.3. Point 4 above would then cause bias to an OR of 1.23 under our conservative assumption.
Another source of a (slight) dilution of the effect estimates was the fact that we were unable to link the national stroke register to the Scanian population, which means that a small fraction of the controls may have been cases.
The correlation between modelled NO
x and experienced disturbance from air pollution was high in the public health survey, which strengthens the validity of our residential exposure assessments.[
34]The proportion that was disturbed by air pollution near their residence was 53% in the highest NO
x-category (≥30 μg/m
3), 36% in the NO
x-category 20–<30 μg/m
3, 18% in the NO
x-category 10–<20 μg/m
3 and 14% in the lowest NO
x-category (<10 μg/m
3).
The biased sampling analysis, where differences in coverage between the hospital admission areas were accounted for, indicated that difference in coverage between hospitals did not substantially influence the effect estimates in this study, an ÔR
1;Party_adj
of 0.98 (0.94–1.02).
For the two-phase analysis, the participants in the public health survey of 2004 provided a sample of potential controls with an already available data set on important confounding factors. There was selective participation in the public health survey with respect to income, age, education, sex, marital status and country of birth. There was also a group-level negative correlation between degree of participation and air pollution, which suggests a bias away from the null. The second-phase analysis suggests an overestimation of the effect (although not statistically significant): the partly adjusted ÔR
2;unadj
for a 10 μg/m3 increase in NOx was 1.09 (0.95–1.26) whereas the partly adjusted first-phase ÔR
1;Party_adj
was 0.99 (0.95–1.02). We note that using only second-phase data would in this study perhaps have led to false conclusions due to the selective participation of the second-phase controls and that the two-phase design is a strength of this study.
The indication of a protective effect by NO
x in rural areas, OR = 0.88 (0.82–0.94) is hard to explain apart from it might be a chance finding. We do not see evidence in our data for residual confounding or biased exposure estimates being plausible explanations for this seemingly protective effect. When adjusting for three of the most important risk factors for stroke (diabetes, smoking, hypertension), the effect estimates change only slightly (Table
3). We find it unlikely that strong residual confounding remains only in the rural areas. In addition, it is also unlikely that exposure measurement errors would cause bias below null, since reversal of the direction requires both that the variance of the error is larger than the variance of the true error and a strong negative correlation between the error and the true value.[
25]
The results in relation to previous studies
Long-term effects on stroke risk have been observed previously in areas where levels of air pollution were considerably higher than in this study. In the UK, an increase in stroke-related mortality of 37% was observed between the lowest and highest quintile group of modelled NO
x (mean 47.6 μg/m
3 and 61.9 μg/m
3, respectively).[
10] Miller and colleagues from USA reported a hazard ratio for first-ever cerebrovascular events (where cerebrovascular deaths were included) in women of 1.35 associated with a 10 μg/m
3 increase in PM
2.5, where the mean value of PM
2.5 was 13.5 μg/m
3.[
11] This corresponds to higher levels of air pollution than in our study area. The hazard ratio for cerebrovascular deaths, which consisted of 20% of the cerebrovascular events, was 1.83 (1.11–3.00). Miller and colleagues recorded first-ever events, as we do in this study, and used monitoring sites to asses exposure. In the Norwegian study, which found an OR for cerebrovascular mortality of 1.04 (0.94–1.15) associated with a 10 μg/m
3 increase in modelled NO
x, the levels of NO
x were rather similar to the levels in Scania, although the composition of air pollution has most likely changed substantially between the 1970s (Norwegian study) and the 2000s (our study).
Perhaps other factors, similar in the Norwegian and Swedish populations, account for the similar results of the two studies. One of those factors might be the rather low concentrations of air pollution, perhaps yielding a larger influence from potential exposure measurement error, though in the Norwegian study, other outcomes showed clear associations with air pollution concentrations.[
12]
A Swedish study on myocardial infarction and long-term exposure to air pollution indicated an effect on out-of-hospital deaths but not on non-fatal cases.[
35] We could not include out-of-hospital deaths in this study, and we restricted our cases to first-ever strokes. The proportion of out-of-hospital deaths seems small in our population (1% in one of the hospital admission areas), the group of male stroke patients with mild strokes who were sent home without being admitted to the hospital was larger (8%). We cannot rule out a potential effect in those groups of stroke cases, or in the rather small fraction of stroke cases that for some reason were not registered in the regional stroke register. However, we see no tendency for a more pronounced effect in cases with a short survival after stroke (30 days or less), with a first-phase OR of 0.96 (0.82–1.13) associated with a 10 μg/m
3 increase in NO
x.
The mean age at diagnosis of the study subjects in this study was rather low (69 years). We decided to include only individuals who were born between 1923 and 1965, since in an older population, competing risks are a problem when making inferences about exposure. Also, the second-phase control material was restricted to individuals born 1923 and later and thus, we did not have the opportunity to adjust for important risk factors in a population born before 1923. Without any age- or first-time stroke restrictions, the mean age in our case group was 78 years (range 3–103 years). Studies on air pollution and stroke often lack an upper age restriction, especially for short-term exposure,[
3‐
6,
8] but also for long-term exposure.[
9,
10] The associations observed in many studies on short-term exposure to air pollution may be partly attributable to the (older) population studied. The Norwegian study (which did not observe an association between long-term exposure to air pollution and cerebrovascular mortality) followed males aged 42–75 years.[
12] The USA (positive) study included women aged 50–79 years.[
11] In our study, we observe no statistically significant modification of effect by sex or age.
Thus, there seem to be evidence for an effect by long-term exposure to air pollution on cerebrovascular mortality, where air pollution concentrations are higher than in our study area. To our knowledge, this study is the first to explore hospital admissions for ischemic stroke in association with long-term exposure to air pollution. The only other study on long-term exposure to air pollution and cerebrovascular events also includes cerebrovascular mortality.[
11] In that study the hazard ratio decreased in the mixed group (cerebrovascular events and cerebrovascular deaths: Hazard ratio = 1.35) compared to the group with only cerebrovascular deaths (Hazard ratio = 1.83). The evidence is therefore less clear for cerebrovascular hospital admissions than for cerebrovascular mortality.
Issues for future research
Using the second-phase data, we observed an effect modification by smoking status, which suggested an effect by an increase of 10 μg/m3 in NOxon ischemic stroke risk for non-smokers but not for smokers. Converted to the two-phase estimate, the increase in risk for non-smokers was no longer present (OR1+2 = 1.00). However, the idea that smokers would be less sensitive to air pollution than non-smokers is interesting, and should be further explored.
In a negative study, it is of great importance to rule out potential sources of error that might lead to bias towards the null. We concluded that although our exposure assessment was of very high quality, exposure misclassification might have biased our results towards the null. Adjusting adequately for socio-economic factors might reduce that bias. Researchers in low-level areas aiming at establishing whether an association between long-term exposure to air pollution and ischemic stroke risk have a challenge in reducing misclassification of exposure.