We only observed J-shaped association between 15-day average temperature and respiratory mortality in the warm period; the other associations did not diverge from linearity. Immediate heat effects could be seen on every outcome in the warm period, even on respiratory mortality in the cold period; while in the cold period, prolonged cold effects could be seen on every outcome except for respiratory mortality. Previous studies also found immediate heat effects and delayed cold effects [24–26]. The strongest immediate heat effect in the warm period was found in association with respiratory mortality, stronger for the whole population than for elderly people (65 years and older). However, the prolonged heat effect on respiratory mortality, as well as the immediate heat effects on cardiovascular, cerebrovascular and cardiorespiratory mortality in the warm period appeared with similar magnitudes for the two age groups. The strongest cold effect in the cold period was found in association with ischemic heart diseases mortality, with similar effect magnitudes for the whole as well as the elderly population. The prolonged cold effects on cardiovascular, cerebrovascular and cardiorespiratory mortality in the cold period appeared also with similar magnitudes for the two age groups. When considering PM2.5 or UFP with lag 2 as confounders, there were no relevant changes for two-day average temperature effects, and a drop in 15-day average temperature effects on cardiovascular and cardiorespiratory mortalities
For the J-shaped exposure-response function, we found 21.3°C as our most appropriate threshold temperature. Curriero et al.  got 19°C to 21°C as "minimum mortality temperature" for New York (NY), Philadelphia (PA), Baltimore (MD), Washington, D.C., U.S., while Ballester et al.  got 22°C to 25°C for Valencia, Spain. Those cities are all located on similar latitudes (from 38°54'N to 40°54'N) as Beijing (39°54'N), and "turning-point temperatures" were all close to each other. However, both authors observed V-shaped temperature-mortality functions.
For the convenience of comparison, we re-calculated the percentages based on a 1°C increase in lag 0-1 average temperature in the warm period, which resulted in 2.5% and 1.9% increases in respiratory and cardiovascular mortality, respectively. A study conducted in four Asian cities including Beijing reported by Chung et al.  found much higher threshold temperature (31°C) and temperature effect estimates (10.5% and 7.6% per 1°C increase, respectively). The fact that they used daily apparent temperature and also included the entire Beijing (including the suburban area with approximately 4,100,000 inhabitants, the fifth national census in 2000, http://www.stats.gov.cn/tjsj/ndsj/renkoupucha/2000pucha/pucha.htm) as their study area might be the reasons for the differences. However, Almeida et al.  reported estimates which are more comparable to ours in their study in Lisbon (38°42'N), Portugal (1.7% and 2.4% per 1°C increase, respectively) using also daily apparent temperature. There are also studies pointing out heat effect on ischemic mortality [26, 29], which has not been found in our results.
Similar to the present study, several authors [7, 27, 30, 31] observed greater effect on respiratory mortality than on cardiovascular mortality. In our dataset, daily death counts due to chronic lower respiratory diseases accounted for approximately half of respiratory diseases. However, our explorative analysis revealed that the magnitude of effect on chronic lower respiratory diseases mortality was as high as 98% of respiratory diseases mortality (data not shown). This may reflect that health status of people suffering from chronic respiratory diseases rapidly deteriorates during hot periods , which should be kept in mind and considered as priority when setting up preventive strategies during heat events.
After an analysis within the warm period, we found that the order (April to September in 2003 is the 1st warm period, April to September in 2004 is the 2nd one, and April to August in 2005 is the 3rd one.) showed no significant interaction with two-day average temperature, and therefore indicates that there was no heat effect modification by potential population adaption to heat or by possible increasing prevalence of air-conditioning year by year.
After re-calculation of the cold effects in the cold period, we found 1.1% and 2.3% increases in cardiovascular and ischemic heart disease mortality of our whole study population associated with 1°C decrease in 15-day average temperature, respectively. Analitis et al.  reported a higher estimate (a 1°C decrease induced a 1.7% increase in cardiovascular mortality in cold seasons) within the PHEWE project, which might be attributed to their use of a 16-day average of minimum apparent temperature. Moreover, the PHEWE project included very cold northern cities such as Helsinki and Stockholm. However, the Eurowinter study  found that people in cold regions such as Finland did not experience more winter excess mortality than people in mild regions such as London; Donaldson et al.  observed no excess ischemic heart disease mortality as temperature fell from 10.2°C to -48.2°C in Yakutsk, eastern Siberia. Both findings reflect the possibility of population acclimatization to climate and maybe also to future climate changes. However, the associations of mortality with environmental temperatures are also strongly modified by behavioural and social factors (e.g. clothing, housing conditions) . In Beijing, the residential heating system works every year from November 15th to next March 15th, regardless of outside temperatures. However, we found no interaction between heating and 15-day average temperature after an analysis within the cold period, showing that residential heating didn't modify the cold effect. This might also reflect that the study population exposed themselves to outdoor temperature although they probably spent a lot of time indoors.
Our study showed effects of increasing temperature on respiratory mortality even during cold season. This is contrary to our initial hypothesis, although the same situation has been observed by Kunst et al.  in The Netherlands. We therefore investigated the exposure-response functions between 2-day or 15-day average temperature and mortality due to influenza and pneumonia (J10-J18) and chronic lower respiratory diseases (J40-J47) (data not shown). Interestingly, we observed different effects regarding the two mortality categories. Whereas a decrease in temperature was associated with an increase in mortality due to influenza and pneumonia (as expected), we found opposite effects for mortality due to chronic lower respiratory diseases. In a previous study, Hampel et al.  have reported differences in the associations between a temperature decrease and several blood markers of inflammation and coagulation in patients with coronary diseases and patients with pulmonary diseases. They hypothesized that there might be different disease patterns as well as patient characteristics and medication responsible for the observed differences in the effects. Nevertheless, although we have no hint of a higher misdiagnosis for respiratory deaths than for deaths due to other causes, we cannot rule out this possibility.
Some studies [36, 37] have shown that respiratory mortality increases more for individuals of 65 years and older compared to the general population when air temperature increases. One possible explanation is that aged people, especially COPD patients, are likely to have bad excess heat dissipation through circulatory adjustment. The heat stress increases their risk of developing pulmonary vascular resistance secondary to peripheral pooling of blood or hypovolemia . However, our results didn't show a higher risk for respiratory mortality in individuals of 65 years and older, although half of the respiratory deaths in our study period were due to chronic lower respiratory diseases (J40-J47, mainly COPD, data not shown). It can be speculated that aged people in Beijing pay more attention and expose themselves less to heat. Possible mechanisms through which high temperature increases cardiovascular mortality include enlarged skin vessels and facilitated sweat, leading to falling blood pressure, increased cardiac work load and loss of fluid and salt, further leading to haemoconcentration , a "thrombosis promoting" state. The activation of coagulation and inhibition of fibrinolysis lead to diffuse microvascular thrombosis. Besides, heat-induced release of interleukin (IL)-1 or IL-6 into systemic circulation results in damage and hyperactivation of endothelial cells.
When temperature decreases, the cold receptors in skin are stimulated, the sympathetic nervous system regulates the catecholamine level to increases  and then the skin vessels constrict to reduce heat loss. Blood pressure increases consequently, and approximately 1l of blood plasma is shifted from skin and legs to central body parts, then removed by urine or shifted to extra-cellular space. The shift of blood plasma leads to haemoconcentration, then the concentrations of red and white blood cells, platelets, certain clotting factors, cholesterol and fibrinogen, as well as blood viscosity all go up, promoting clotting and thrombosis. Moreover, protein C, which is an anticoagulant, moves out to the extra-cellular space with blood plasma. The rise of blood pressure may lead to oxygen deficiency in the cardiac muscle which might induce myocardial ischemia or arrhythmias. If the rise of blood pressure is too sudden, there is the possibility of vascular spasm and a rupture of an atherosclerotic plaque that induces a thrombus [39, 41–43].
Strengths and limitations
This study was based on a population as large as seven million inhabitants, among which the daily cardiorespiratory death count reached 40. This ensured the statistical power of the analysis. Moreover, we did sensitivity analyses by including PM2.5 or UFP concentration levels as confounders. Both PM2.5 and UFP  have been shown to be associated with mortality. As Beijing is known as one of the most polluted cities of the world, controlling for these two air pollutants was an important strength of the present study. Some other studies [36, 45, 46] also considered PM2.5, PM10 or black smoke as confounders.
However, there are also limitations of the present study. Firstly, we got both, the meteorological and the air pollution data from only one monitoring station, which may lead to misclassification of the exposure level. This misclassification is non-differential and should bias the effect estimates towards the null. However, further data on daily meteorological parameters from an internet service (Weather Underground 2011) was obtained for a station located in the center of Beijing. Data from the two sources showed a good agreement (Pearson correlation coefficient >0.99 for air temperature). Secondly, ozone is a potentially important confounder to heat effect, but we had no such data for a sensitivity analysis.