Seasonality of mortality and, in general, of disease is a well-known phenomenon in many regions and countries worldwide. Numerous studies have been conducted in industrialized countries of the mid-latitudes [1–7] that relate a multitude of causes of death to seasonal incidence (e.g., cardio-respiratory diseases, infectious disease) [5, 8–13]. Likewise, seasonal fluctuations have been observed for tropical climates [14–18], despite the less-pronounced intra-annual climatological variation that is mainly related to seasonal differences in precipitation. Nevertheless, the number of studies focusing on tropical climates is limited, especially for Asian countries.

Although seasonal variations are to some extend driven by seasonal variations in weather, they underlie various non-atmospheric influences. These influences have fundamentally modified the shape of the seasonal pattern over recent centuries [5, 19–22]. In developed countries, a shift from a summer peak in mortality towards a winter peak has been observed (ibid). In contrast, tropical countries have been, and still are, associated with excess summer mortality; this is often explained by a high prevalence of infectious and diarrhoeal disease [17, 18, 23–25]. The modifying effect of non-atmospheric parameters is well-demonstrated by the existence of different seasonality regimes within the same climatic region. For instance, major differences between urban and rural areas could be observed in France in the 18th century [26], and were also found between White and Afro-American groups in Philadelphia in the same century [27]. In a more recent study, education serving as a proxy for socioeconomic status has been highlighted as a determinant for seasonal fluctuations of mortality [5].

Apart from climate, cultural and behavioural aspects apparently play a major role in shaping the seasonal distribution of mortality [5, 28, 29]. Paradoxically, studies show that countries with a relatively warm or mild winter climate, such as Spain, Portugal, Italy, or the UK and Ireland, experience much greater excess winter mortality than countries with harsh climatic conditions during winter, such as Finland, Norway, or Russia (Siberia) (ibid). Better adjustment and social adaptation to the cold in countries with cold winter climates have been cited as explanations (ibid). There is evidence that with the same outdoor temperatures, people living in colder climates wear warmer clothes and protect themselves better against the cold [29–31].

Understanding the impact of seasonally varying factors, the effects of atmospheric conditions and the modifying effect of non-atmospheric influences can make a contribution to establishing effective public health measures. Whereas in a recent study [32] we investigated the short-term effects of thermal conditions, this study seeks to asses mid-to long-term seasonal effects and atmospheric influences. To date, few studies have investigated the atmosphere-mortality relationship in a tropical developing country; whereas the vast majority of all such research has concentrated on an assessment of seasonal effects rather than immediate meteorological effects. In focusing on seasonality we were able to set the findings of this work in the context of other studies thus allowing comparison of results. Moreover, conducting a multi-stratified analysis enabled us to reach a better understanding of the various non-atmospheric effect modifiers. In particular, we focused on the differences between urban and rural areas, gender differences and differences between regions with different socioeconomic status (SES). We also considered various causes of death and age-specific effects.

Mortality data were provided by the Bangladesh Bureau of Statistics (BBS) for the period from 2002 to 2007. The data were collected within the Sample Vital Registration System (SVRS), which has existed in its current version since 2002. The SVRS comprises and surveys 1,000 Primary Sample Units (PSU) in rural and urban areas, and in the statistical metropolitan area (SMA). A PSU is a compact cluster with approximately 200 households with a household size of 4-5 members (4.7 on average). A number of 640 PSUs is located in rural areas, comprising 132,646 households; 280 PSUs are located in urban areas with 57,852 households and 80 PSU are placed in the SMA with 16,024 households. The SVRS provides information on housing, household, and population characteristics that are continuously updated. Data is collected using a dual recording system. Initially, vital events are collected by a locally-recruited recorder (System 1). Subsequently, data are collected in retrospective by a group of officials from the BBS on a quarterly basis (System 2). Afterwards, data are matched by pre-designed matching criteria and classified into matched, partially-matched and non-matched events. Partially-matched and non-matched events are subject to further verification through field visits. The following information about each fatal event was recorded: name of the deceased, date of birth, date of death and sex of the deceased. Moreover, a cause of death is attributed, which, however, is not medically certified. For further information see [33].

For this study all accidental deaths, maternity-related deaths, data from the statistical metropolitan area, and the data of 2002 were excluded. Furthermore, we excluded deaths of infants younger than one year, as births exhibit seasonal variations which could confound our analysis. As severe flooding submerged major parts of the country in 2004, we conducted a sensitivity analysis running our analysis with and without data from this year. Results were mostly unaffected, except for diarrhoeal mortality and other-cause mortality. Therefore, the analysis on those two causes of death was done excluding deaths occurring in 2004. In total, 21,551 deaths were analyzed. Stratified-time series plots of the daily death counts are provided in an additional file (Additional file 1).

Daily mortality was defined as daily crude death rates (number of deaths per population). In order to facilitate comparison between different subcategories we calculated the percent deviation from average mortality for each day of the year from 2003 to 2007. Average daily mortality (expected mortality) was calculated separately for each year in order to account for inter-annual changes in mortality.

In order to assess the extent of seasonality and its age-specific characteristics, we calculated a seasonality index (Φ) for each year of life, defined as the ratio of the number of deaths in the months with the highest and lowest mortality (see Formula 1).

(1)

With Φ being the seasonality index, N_Max and N_Min the numbers of deaths in the months with the highest and lowest death counts in year i, and n the number of years from 2003 to 2007.

To display seasonal variations of mortality and age-specific seasonality of mortality we smoothed the data with penalized spline (P-spline) smoothers. Unlike many other seasonality studies we chose a non-parametric smoothing approach in order to account for multi-modal seasonal distributions. Data analysis and smoothing was carried out using R (Version 2.11.0) and the R package 'mgcv'. P-splines applied in this analysis are as proposed by Eilers and Marx [34].

The analysis of seasonality by SES was conducted on the administrative level of Zilas, which are subunits of the six divisions existing in Bangladesh. To distinguish between Zilas with high and low SES, we derived four socioeconomic factors by conducting a factor analysis (Varimax rotation) using R (Version 2.7.2). Variables considered for factor analysis were child mortality rate, child/woman ratio, literacy rate, fertility rate, source of drinking and non-drinking water, infant mortality, insolvency rate and use of solid fuels. At least three of the four derived factors being below or above the 50th percentile were criteria for Zilas to be categorized as either low or high SES. Of the 64 Zilas, 23 were categorized as having poor SES and 25 were categorized as having high SES. The remaining Zilas were treated as reference. Like in most developing countries, differences between sub-populations are strongly pronounced, but also differences between the considered divisions were marked. The relative standard deviation for the use of surface water as drinking water reached up to 400%; also for infant and child mortality or insolvency the relative standard deviations ranged between 30 and 70%.

In this study we identified distinct seasonal variations of mortality, with the shape of the seasonal pattern depending on the cause of death, age, gender, and location, as well as SES. Unlike findings from other tropical regions [17, 18, 23–25, 35–39], we did not find a primary maximum in mortality during the hot and humid/rainy season. In our data the summer excess mortality may have been represented by the secondary maximum from April to July. However, the main peak in mortality occurred during the cold season. This observation suggests that Bangladesh is undergoing an epidemiological transition as the incidence of diarrhoeal deaths reduces considerably and cardio-respiratory diseases become more dominant. A transition in the seasonal pattern - from a summer to a winter peak - has also been described for Europe and Japan, and SES changes and improved medical care served as explanations [5, 21]. Indeed, a study analysing death counts from 1972-1974 from a rural area in Bangladesh found a winter maximum for dysentery and chronic diarrhoeal mortality. The author ascribed this pattern to the high quality of diarrhoeal health care in that particular (intervention) area and acknowledged that such a mortality pattern is possibly not found in other areas [14]. In a follow-up study, in which data from 1982 to 1990 was analyzed, a winter maximum was found for all-cause mortality; however, diarrhoeal deaths peaked during the hot and humid season [15]. To date, the only lower-latitude regions exhibiting a mortality peak during the cold season are Mexico City (Mexico) and São Paulo (Brazil), both located within higher altitudes and exhibiting temperate climatic conditions (type C-climates according the Köppen-Geiger classification) as well as a generally higher SES [16]. In addition, a higher mortality during the cold months was found for under-five mortality in Nairobi (Kenya), also classified as a temperate type C-climate [40, 41]. To our knowledge, the only country with a tropical type A-climate for which a winter excess mortality has been established is Bangladesh.

Nowadays, respiratory and cardiovascular diseases are the leading causes of death in Bangladesh. We found that both diseases as well as all-cause mortality peaked during winter, despite the perception of the relatively cold season as being thermally comfortable. These findings suggest that winter excess mortality is not a consequence of seasonal low absolute temperatures but, rather, is a consequence of a seasonal fall in average temperature. The adverse effect of relatively low temperatures was also demonstrated in a recent study using the same data set [32]. Regression models adjusted for the confounding effects of trend, season and day of the month revealed an increase in mortality with temperatures falling below a threshold situated in the 90^th to 95^th percentile of the temperature distribution. While adequately adapted to heat, the Bangladeshi population seems less prepared for the winter period. The importance of relative cold is also highlighted by research conducted by Douglas et al. [1] and the Eurowinter Group [42].

The pronounced summer peak of all-cause and cardiovascular mortality in urban areas is possibly related to the urban excess (equivalent) temperature, the so-called urban heat island. Excess temperatures amounting to several Kelvin in monthly averages were found in urban areas of (sub)tropical regions [43, 44]. Nevertheless, a higher susceptibility of the urban population towards heat might also be an explanation. A higher susceptibility towards heat might also be the cause of the secondary summer maximum of cardiovascular mortality in regions with high SES.

In this study the complexity and dynamic nature of diarrhoeal disease pathogenesis is once more illustrated. Several agents that induce diarrhoeal disease favour different meteorological and hydrological conditions for their replication and survival. High temperatures generally promote the growth of bacteria, whereas, for instance, shigella epidemics show a high incidence rate during the cold season [45]. Furthermore, there is evidence that moderate temperatures favour certain agents, such as the rotavirus [46, 47]. Additionally, hydrological aspects, such as stagnant water, a lack of dilution, contamination, or the break down of water systems associated with no, high, or ongoing rainfall, determine the spread of diarrhoeal pathogens [46, 48]. Cholera incidence showed a bimodal distribution in Bengal that seemed to be driven by hydroclimatological factors [49, 50]. Low precipitation and low river discharge during spring was associated with the first outbreaks of cholera in Bangladesh whereas high precipitation and peak streamflow of rivers during the monsoon season was associated with the second peak during the monsoon season (ibid).

Given this background information, the interpretation of our findings is complex. Nevertheless, we suggest that the summer peak in diarrhoeal mortality is predominantly caused by bacterial contamination during the hot season favouring replication of agents, whilst the post-monsoonal peak probably results from overstrained sewage systems, flooding and stagnant water. The pronounced seasonal pattern in rural areas and areas with low SES is possibly due to reduced coping abilities in these regions. The absence of a summer peak in urban areas and the reduced post-monsoonal peak might be due to improved access to fresh aliments, cooling systems as well as more developed drainage and sewage systems. Regarding female diarrhoeal mortality, the nonexistence of seasonal fluctuations is surprising as females are facing an increased risk of dying from diarrhoeal mortality. Women are usually bound to their homes consuming home made food, whilst men are commonly pursuing an income generating activity outside that forces them to eat food either self-carried or purchased at some sort of restaurant. Particularly, during the summer season the risk of bacterial contamination is thus increased. In areas with high SES, higher sanitary standards, improved water and drainage systems and higher levels of education might have resulted in reduced diarrhoeal seasonality.

Our analysis further showed that seasonality is strongly age-dependent. Children in their first years of life exhibited a high magnitude of seasonality, demonstrating their susceptibility towards environmental conditions. This susceptibility to season and environment was even more pronounced in rural areas, probably resulting from poor primary health services as well as poor living standards. Although young and middle-aged adults are likely to be most excessively exposed to environmental influences they were the only group to show weak or no seasonality. The increasing seasonality index with increasing age reflects the growing susceptibility of the human organism towards environmental conditions. The observed decrease in seasonality between the ages of 80 and 90 might be due to the so called selection effect, described in detail by Rau [5]. Briefly, the phenomenon is the consequence of a heterogeneous population in which frail individuals face higher mortality, leaving a robust subset. This robust subset is less susceptible to seasonal effects, lowering seasonality up to a certain age beyond which individuals once again become more susceptible.

This study demonstrated the importance of considering seasonal impacts on mortality and revealed several target groups and diseases for which the consideration of seasonality seems particularly crucial. We showed that the effect of season or seasonally changing environmental conditions depends on preconditions in a subpopulation or region. In particular, rural areas showed a high magnitude of infant and child seasonality, while urban areas were strongly associated with summer excess mortality, especially for cardiovascular mortality. Children and elderly people were affected most by seasonal effects. The consideration of such seasonal effects could help to place public health interventions most effectively. Generally, protection measures against cold or heat among infants or those under hospital care can help to avoid or overcome critical states of health. We emphasise that knowledge on the interaction between seasons, atmosphere, and health is especially necessary in times of climate change so that its possible impacts can be mitigated.