The study area, Scania, is the southernmost region in Sweden. It covers 11 350 km2
, which is approximately 2% of the Swedish land area.[18
] In December 2008, Scania had around 1.2 million habitants, approximately 13% of the Swedish population.[19
] Levels of pollutants in Scania are generally higher than in Sweden overall, but rather low from an international point of view.[17
The study was based on a series of 14 770 hospital admitted cases of stroke that occurred between January 1, 2001 and December 31, 2005 in the region of Scania, which we obtained data on from Riks-stroke, the national quality register for acute stroke (national stroke register). This register has been described in detail elsewhere[20
] and has an estimated coverage of hospital admitted ischemic stroke cases of around 90%. However, the register does not include stroke patients who die out of hospital before the stroke is registered. A study in one of the hospital admission areas in Scania suggests that 1% of the stroke patients had a fatal stroke before a medical examination and thus were not included in the register.[22
] Moreover, that study showed that a subgroup of male patients with mild strokes were to a larger extent sent home without being registered (8% of the total patients). To obtain as high quality a case definition as possible, we chose not to incorporate information from the cause-of-death register. The outcome in this study is therefore ischemic stroke hospital admissions.
First-time episodes of ischemic stroke in persons born between 1923 and 1965 were included, resulting in 4 904 first-phase cases (Table and Figure ). The median age at diagnosis was 72 years.
First-phase descriptive data for the variables birth year category, sex, birth country and marital status.
Flow-chart for cases and controls.
The proportion of computer tomographies among our stroke cases was 99% for ischemic stroke cases and 97% in total, which indicates an accurate classification of types of stroke.
Out of the 4904 first-phase cases, 1552 (32%) were registered as deceased the 29th of June 2009. Out of the 1552 deceased, 152 (10% of the deceased, 3% of the total number of cases) died within 10 days of the day the stroke occurred, 244 (16%, 5%) died within 30 days after the stroke and 438 (28%, 9%) died within 6 months after the stroke.
The national stroke register provided information on smoking status, diabetes, and medication for hypertension for 4 375 of the 4 904 ischemic stroke cases (Table , Figure ). Statistics Sweden, which is a central government authority for registration and provision of official statistics, provided information on the marital status and country of birth of all cases.
Second-phase descriptive data for the variables diabetes, smoking, medication for hypertension.
Information on year of birth, sex, country of birth, marital status, and location of residence was available from Statistics Sweden. The information was dated December 31, 2002. Individuals born between 1923 and 1965 who were residing in Scania were included, which resulted in 556 912 controls (Table , Figure ).
We stratified the controls by exposure (NOx above or below 20 μg/m3) to see if potential risk factors were associated with exposure level. This stratification suggested that country of birth and marital status might confound unadjusted estimates of exposure effect (Table ).
Information on smoking status, diabetes, and medication for hypertension was obtained for controls that had participated in a large public health survey distributed as a mailed questionnaire in 2004. The target population of the survey was individuals between 18 and 80 years of age who were residing in Scania on June 30, 2004. The recruiting was sampled on geographical area. In total, 27 963 people responded to the questionnaire, which corresponds to an overall participation rate of 59%. There was selective participation in the public health survey with respect to income, age, education, sex, marital status and country of birth. The most pronounced difference in participation was between people born outside Europe (33.9%) and people born in Sweden (59.5%).[23
] A group-level negative correlation between degree of participation and air pollution was present, but data did not allow investigating if it was also present on the individual level.
We matched each ischemic stroke case to two controls (or if two were not available, to one control) from the public health survey. We matched on birth year category, sex, and on the geographical areas used in the public health survey, resulting in a total of 5 333 controls (Figure ). Of these controls, 4 716 had complete information on smoking status, diabetes, and medication for hypertension (Table , Figure ).
Assessment of individual levels of exposure to air pollution is challenging,[24
] but modelling of air pollution using geographic information systems (GIS) together with information on emissions, meteorology, and traffic intensity enables one to model levels of air pollution with high resolution, in terms of both geography and time.
As a marker of exposure to air pollution, we chose NOx
, which is considered to be a good proxy for air pollution in general, especially for substances generated by traffic.[2
] We used a GIS linked to an emissions database with approximately 24 000 sources of emissions to model individual residential levels of NOx
The modelled concentrations were validated against measurements from 23 monitoring sites in Scania, yielding an overall Pearson correlation coefficient with the modelled values of r = 0.69.[18
] The correlation between the measured concentrations and the error (difference between measured and modelled concentration) was r = -0,016 and between the modelled concentrations and the error r = 0.55, suggesting a classical measurement error structure.[25
] The standard deviation of the measured values was 7.7 μg/m3
and of the modelled values 8.3 μg/m3
The modelled annual means of NOx
in Scania are presented in Figure . We modelled NOx
on an hourly basis in a grid with a spatial resolution of 500 × 500 m. [Susanna Gustafsson, A geographical and temporal high resolution database for dispersion modelling of environmental NOx
in Southern Sweden, submitted] A previous study indicated that a resolution of 200–400 m is suitable in urban areas, but the larger size of our study made 500 × 500 m a feasible choice in spatial resolution.[26
] Also, in an ongoing study, measured data is compared with modelled data using different spatial resolutions. Preliminary validation results from that study indicate that a resolution of 500 × 500 m and a resolution of 250 × 250 m yield rather similar results [Emilie Stroh, Personal communication]
The distribution of the modelled NOx concentrations in Scania. The mean NOx concentration for the cases was 12.9 μg/m3, with a standard deviation of 7.8, and for the controls 12.5 μg/m3, with a standard deviation of 8.0.
For the cases, the individual geocoded residential addresses on December 31 of the year before diagnosis of ischemic stroke were linked to exposure data, and exposure was assessed as the annual mean (at that address) before the date the stroke occurred. The first-phase controls were linked to the address they had on December 31, 2002. The second-phase controls were linked to the address they had in the public health survey of 2004. The controls were randomly distributed over the days of the study period (2001–2005), with the same distribution as the ischemic stroke cases, and their annual mean exposure was assessed as the annual mean before that randomly assigned date.
Although our aim was to investigate long-term exposure to air pollution, the data allowed assessing exposure only one year back in time for the majority of our study subjects. However, the geographical contrasts in exposure over time are fairly constant. We therefore consider the annual mean of NOx the year before the stroke-date as a good proxy for NOx exposure over a longer time period.
A large proportion (78%) of the second-phase controls had not changed residential address during the past 10 years according to their answers in the public health survey. For the first-phase controls, we had no information on residential address back in time. Register data on residential address back in time was available for roughly half of the cases (based on what part of Scania they lived in). Although not entirely exact, that data indicated that the proportion of cases that had not changed residential address 10 years previous to the stroke was similar to the proportion of the second-phase controls (78%). Out of those we had information on, around 5% had changed residential address the year previous to the stroke year.
The study was designed as a two-phase study, a design that can increase precision and reduce bias from selective participation in case-control studies.[27
] In this study, the two-phase design was used to investigate whether residual confounding of the variables smoking, diabetes, and medication for hypertension occurred. These variables were only known for a subsample of the study subjects (the second-phase subjects), for 4 375 of the ischemic stroke cases and 4 716 of the controls (Table , Figure ).
In a two-phase design, the log-odds ratio log(ÔR1+2
) for the combined first- and second-phase data (1+2) can be estimated with adjustments for confounders in the following way.[27
Logistic regression analysis including the variable birth year category, sex, marital status, and country of birth as covariates yielded ORs that were denoted as the "partly adjusted ORs". This corresponds to ÔR1;unadj and ÔR2;unadj in formula 1 for the first- and second-phase subjects, respectively, hereafter denoted ÔR1;Party_adj and ÔR2;Party_adj. The second-phase OR, adjusted also for smoking, diabetes, and medication for hypertension was denoted ÔR2;adj in formula 1 or as the "fully adjusted OR".
The NOx variable was analyzed both as a continuous measure and categorized as: 0–<10, 10–<20, 20–<30 and >30 μg/m3, with 0–<10 μg/m3 as reference category. The p-value for the effect of the continuous NOx variable was used as a p-value for trend. The variance expression in formula 1 has been derived only for categorical variables; therefore we refrained from calculating the two-phase confidence intervals for the continuous NOx variable.
For the partly adjusted first-phase analysis, we explored potential effect modifications on the association between the continuous NOx-variable and ischemic stroke for the variables birth year category (> 1940 or ≤ 1940), sex, birth country and residing in urban/rural area. For the fully adjusted second-phase analysis, we explored effect modification by smoking status.
In order to take account for the different registration coverage between the hospitals, we applied a sampling technique for the controls in which they were assumed to be recruited with a probability according to the registration coverage in their hospital admission area. The data were analyzed as described by Weinberg and Wacholder,[29
] (page 967–969), where an offset is added to the model to account for the "biased" sampling. The offset is calculated for each individual according to the probability of recruitment.
All analyses were done using SAS version 9.1 (SAS Institute Inc., Cary, NC).