Sample
In the present analysis, all cases (n = 25,499) of coronary deaths and non-fatal AMI aged 25–74 years recorded in the MONICA/KORA Myocardial Infarction Registry between 1 January 1985 and 31 October 2010 were included. The population-based myocardial infarction registry was implemented in 1984 as part of the WHO-MONICA (Monitoring Trends and Determinants in Cardiovascular Disease) project. After the termination of MONICA in 1995, the registry became part of the framework of KORA (Cooperative Health Research in the Region of Augsburg). Since 1984, all cases of coronary deaths and non-fatal AMI of the 25–74 year old inhabitants in the city of Augsburg and the two adjacent counties (about 600,000 inhabitants) have been registered. Data sources for hospitalized patients include 8 hospitals within the study region and 2 hospitals in the adjacent areas. Approximately 80 % of all AMI cases of the study region are treated in the study region’s major hospital, Klinikum Augsburg, a tertiary care centre offering invasive and interventional cardiovascular procedures, as well as heart surgery facilities. Methods of case finding, diagnostic classification of events, and data quality control have been described in more detail elsewhere [10–12]. In brief data on pre-hospital coronary deaths is based on the death certificates collected by the three public health departments within the study region. For patients who are hospitalized with an AMI, a comprehensive set of data on demographics, cardiovascular risk factors, medical history and AMI treatment is being collected by individual interview and chart review.
The study was approved by the ethics committee of the Bavarian Medical Association. All participants with non-fatal AMI submitted written informed consent before being enrolled in the study.
Data collection
AMI survivors were interviewed by trained study nurses during their hospital stays after they have been transferred from the intensive care unit using a standardized questionnaire.
Patients were asked whether they are currently employed (yes/no), have ever smoked or have stopped smoking (current smoker/ex-smoker/never smoked), whether they were diagnosed as having high blood pressure, blood lipids or diabetes prior to the AMI event, and whether they had an AMI before. Self-reported history of hypertension, hyperlipidaemia, diabetes or recurrent AMI (yes/no) was only considered if the chart review confirmed these diseases. Body mass index (BMI) was determined by assessment of height and weight during the hospital stay. Obesity (yes/no) was defined as BMI > 30 kg/m2. Data on medication prior to AMI (antiplatelets, beta-blockers, calcium antagonists, angiotensin-converting enzyme inhibitors, lipid-lowering agents) were collected both from individual patient interviews and chart reviews in AMI survivors. Patients were asked for the exact date and time of symptom onset. This information was validated against the information from the medical chart.
For coronary deaths, information on re-infarction, medication prior to AMI, current occupation, history of hypertension, hyperlipidemia, diabetes, smoking and obesity were requested from the last attending physician. Date and time of hospital admission or death was used as equivalent for date and time of symptom onset.
In addition, air temperature, relative humidity, and barometric pressure were measured on an hourly basis by the Bavarian Air Monitoring Network at one background air monitoring site 7 km south of the Augsburg city center [13]. 24-h mean values were calculated if at least 75 % of the hourly values were available.
Data analysis
Time series model
We used a time series approach to model the association between DST transitions and incidence of non-fatal AMI and coronary deaths. In specific, we applied generalized additive Quasi-Poisson models to accommodate a Poisson distribution with overdispersion for the daily cases of AMI. To assess the influence of DST transition, we included two indicator variables for the week after the spring and the autumn transition. Alternatively, we included indicator variables for the 3 days after the time shifts. As potential confounders, we considered a global time trend, temperature, relative humidity, barometric pressure, and indicators for month of the year, weekday and holidays. Model selection was based on a reduction of the generalized cross-validation criteria (GCV) and the absolute value of the sum of the partial autocorrelation function [14]. For the meteorological factors, we compared the levels of same day, up to four previous days and the average of these five days and chose the term which minimized GCV most. To model nonlinear confounder effects, we used penalized regression splines to optimize the degree of smoothness. The optimal degree was then kept fix to allow a better comparability with sensitivity models. The final model included the following covariates: time trend and previous two day mean relative humidity as regression splines with four and two degrees of freedom, respectively, previous two day mean temperature as a linear term and day of the week as categorical variable.
Excess model
We reduced the time series to the months around the time transition (March and April for the spring shift and September to November for the autumn shift) to construct a prediction model in order to assess the potentially higher rates caused by the time shifts. On that account, we reran the confounder selection for spring and autumn months separately excluding the data of the week following the transition. The optimized spring model included time trend and same day mean relative humidity as regression splines with six and three degrees of freedom, same day mean temperature as a linear term, and month and weekday as categorical variables. The optimized autumn model included time trend and same day mean temperature as linear terms, same day mean relative humidity as regression spline with three degrees of freedom, and month and weekday as categorical variables. We then applied the two regression models to the datasets with weeks following the spring and autumn transitions, respectively, to predict the expected numbers of AMI per day. The incidence rate ratio was assessed as observed over expected events per day and the mean per weekday and corresponding 95 % confidence intervals were calculated.
Effect modification
We conducted stratified analyses of the data sets reduced to spring and autumn months on the basis of the therefore optimized confounder models. In specific, we assessed effect modification by sex, age (≤65 vs. >65 years), working vs. not working, first vs. recurrent AMI, non-fatal vs. fatal AMI, history of diabetes, hypertension, hyperlipidemia, BMI (<=30 kg/m2 vs. >30 kg/m2), smoking status, and intake of the following medications: antiplatelets, beta-blockers, calcium antagonists, ACE inhibitors, and lipid lowering drugs. Since it was not possible to obtain the information on a number of these variables for all patients (e.g. employment status was only requested in the interview and therefore not available for people who died before being interviewed), most of the stratified analyses are restricted to a limited sample size ranging between 89.9 % (diabetes) and 53.3 % (smoking) of the total population.
Sensitivity analysis
To check the robustness of our time series results we conducted several sensitivity analyses. First, we replaced the continuous time trend variable by a fixed effect for the year (S1). Second, we considered the inclusion of a penalized spline for time trend implying an optimization of the degree of smoothness by the model (S2). Third, we additionally modeled previous two day mean temperature and relative humidity by penalized splines (S3). Fourth, we additionally incorporated indicator variables for months and holidays (S4). All statistical analyses were performed with R software, version 3.0.0, package “mgcv”.