Skip to main content
NCBI home page
eLife logoLink to eLife
. 2018 Oct 30;7:e35500. doi: 10.7554/eLife.35500

National and regional seasonal dynamics of all-cause and cause-specific mortality in the USA from 1980 to 2016

Robbie M Parks 1,2, James E Bennett 1,2,3, Kyle J Foreman 1,2,4, Ralf Toumi 5, Majid Ezzati 1,2,3,✉
Editors: Eduardo Franco6, Prabhat Jha7
PMCID: PMC6207431  PMID: 30373715

Abstract

In temperate climates, winter deaths exceed summer ones. However, there is limited information on the timing and the relative magnitudes of maximum and minimum mortality, by local climate, age group, sex and medical cause of death. We used geo-coded mortality data and wavelets to analyse the seasonality of mortality by age group and sex from 1980 to 2016 in the USA and its subnational climatic regions. Death rates in men and women ≥ 45 years peaked in December to February and were lowest in June to August, driven by cardiorespiratory diseases and injuries. In these ages, percent difference in death rates between peak and minimum months did not vary across climate regions, nor changed from 1980 to 2016. Under five years, seasonality of all-cause mortality largely disappeared after the 1990s. In adolescents and young adults, especially in males, death rates peaked in June/July and were lowest in December/January, driven by injury deaths.

Research organism: None

eLife digest

In the USA, more deaths happen in the winter than the summer. But when deaths occur varies greatly by sex, age, cause of death, and possibly region. Seasonal differences in death rates can change over time due to changes in factors that cause disease or affect treatment.

Analyzing the seasonality of deaths can help scientists determine whether interventions to minimize deaths during a certain time of year are needed, or whether existing ones are effective. Scrutinizing seasonal patterns in death over time can also help scientists determine whether large-scale weather or climate changes are affecting the seasonality of death.

Now, Parks et al. show that there are age and sex differences in which times of year most deaths occur. Parks et al. analyzed data on US deaths between 1980 and 2016. While overall deaths in a year were highest in winter and lowest in summer, a greater number of young men died during summer – mainly due to injuries – than during winter. Seasonal differences in deaths among young children have largely disappeared and seasonal differences in the deaths of older children and young adults have become smaller. Deaths among women and men aged 45 or older peaked between December and February – largely caused by respiratory and heart diseases, or injuries. Deaths in this older age group were lowest during the summer months. Death patterns in older people changed little over time. No regional differences were found in seasonal death patterns, despite large climate variation across the USA.

The analysis by Parks et al. suggests public health and medical interventions have been successful in reducing seasonal deaths among many groups. But more needs to be done to address seasonal differences in deaths among older adults. For example, by boosting flu vaccination rates, providing warnings about severe weather and better insulation for homes. Using technology like hands-free communication devices or home visits to help keep vulnerable elderly people connected during the winter months may also help.

Introduction

It is well-established that death rates vary throughout the year, and in temperate climates there tend to be more deaths in winter than in summer (Campbell, 2017; Fowler et al., 2015; Healy, 2003; McKee, 1989). It has therefore been hypothesized that a warmer world may lower winter mortality in temperate climates (Langford and Bentham, 1995; Martens, 1998). In a large country like the USA, which possesses distinct climate regions, the seasonality of mortality may vary geographically, due to geographical variations in mortality, localized weather patterns, and regional differences in adaptation measures such as heating, air conditioning and healthcare (Davis et al., 2004; Braga et al., 2001; Kalkstein, 2013; Medina-Ramón and Schwartz, 2007). The presence and extent of seasonal variation in mortality may also itself change over time (Bobb et al., 2014; Carson et al., 2006; Seretakis et al., 1997; Sheridan et al., 2009).

A thorough understanding of the long-term dynamics of seasonality of mortality, and its geographical and demographic patterns, is needed to identify at-risk groups, plan responses at the present time as well as under changing climate conditions. Although mortality seasonality is well-established, there is limited information on how seasonality, including the timing of minimum and maximum mortality, varies by local climate and how these features have changed over time, especially in relation to age group, sex and medical cause of death (Rau, 2004; Rau et al., 2018).

In this paper, we comprehensively characterize the spatial and temporal patterns of all-cause and cause-specific mortality seasonality in the USA by sex and age group, through the application of wavelet analytical techniques, to over three decades of national mortality data. Wavelets have been used to study the dynamics of weather phenomena (Moy et al., 2002) and infectious diseases (Grenfell et al., 2001). We also used centre of gravity analysis and circular statistics methods to understand the timing of maximum and minimum mortality. In addition, we identify how the percentage difference between death rates in maximum and minimum mortality months has changed over time.

Results

Table 1 presents number of deaths by cause of death and sex. Deaths from cardiorespiratory diseases make up nearly half of all deaths (48.1%), with most deaths in that group from cardiovascular diseases. Next highest during the study period were deaths from cancers (23.2%), followed by injuries (6.8%), with two thirds of those being from unintentional injuries.

Table 1. Number of deaths, by cause of death and sex from 1980 to 2016.

Cause Male Female Total
All cause 43,558,203 42,295,973 85,854,176
Cancers 10,481,582 9,476,530 19,958,112
Cardiorespiratory diseases 20,168,049 21,109,525 41,277,574
Cardiovascular diseases 16,238,344 17,210,556 33,448,900
Chronic respiratory diseases 2,791,652 2,595,950 5,387,602
Respiratory infections 1,138,053 1,303,019 2,441,072
Injuries 4,034,876 1,768,170 5,803,046
Unintentional 2,489,142 1,348,187 3,837,329
Intentional 1,545,734 419,983 1,965,717
Other causes 8,873,696 9,941,748 18,815,444
Open in a new tab

All-cause mortality in males had a 12 month seasonality in all age groups, except ages 35–44 years, for whom there was periodicity at 6 months (Figure 1). In females, there was 12 month seasonality in all groups except 5–14 and 25–34 years (p-values=0.21 and 0.25, respectively) (Figure 2). While seasonality persisted throughout the entire analysis period in older ages, it largely disappeared after the late 1990s in children aged 0–4 years in both sexes and in women aged 15–24 years.

Figure 1. Wavelet power spectra for national time series of all-cause death rates for 1980–2016, by age group for males.

Figure 1.

Open in a new tab

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 2. Wavelet power spectra for national time series of all-cause death rates for 1980–2016, by age group for females.

Figure 2.

Open in a new tab

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Deaths from all causes of death were seasonal in older adults (above 65 or 75 years depending on cause, p-values<0.06) (Figures 1–10 and respective figure supplements), except for intentional injuries and substance use disorders. Deaths from cardiorespiratory diseases, and within it respiratory infections, exhibited seasonality throughout the life-course (p-values<0.06) except for males aged 5–24 years and females aged 15–24 years (p-values>0.11). In addition to older ages, injury deaths were seasonal from childhood through 44 years in women and through 64 years in men (p-values<0.09). Unintentional injuries drove the seasonality of injury deaths for females, whereas both unintentional and intentional injuries were seasonal in males in most ages, with the exception of below 15 years and above 85 years when intentional injuries were not seasonal (Figure 7—figure supplement 1). Consistent seasonality in cancer deaths (Figures 3,4) only appeared after 55 years of age (p-values<0.05). No consistent seasonality was evident in substance use disorders (Figure 9—figure supplement 1 and Figure 10—figure supplement 1) or maternal conditions (Figure 10—figure supplement 2).

Figure 10. Wavelet power spectra for national time series of death rates from causes other than cancers, cardiorespiratory diseases and injuries for 1980–2016, by age group for females.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 10.

Figure 10—figure supplement 1. Wavelet power spectra for national time series of substance use disorder death rates for 1980–2016, by age group for females.

Figure 10—figure supplement 1.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 10—figure supplement 2. Wavelet power spectra for national time series of maternal condition death rates for 1980–2016, by age group for females.

Figure 10—figure supplement 2.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 10—figure supplement 3. Wavelet power spectra for national time series of perinatal condition death rates for 1980–2016, by age group for females.

Figure 10—figure supplement 3.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 10—figure supplement 4. Wavelet power spectra for national time series of endocrine disorder death rates for 1980–2016, by age group for females.

Figure 10—figure supplement 4.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 10—figure supplement 5. Wavelet power spectra for national time series of genitourinary disease death rates for 1980–2016, by age group for females.

Figure 10—figure supplement 5.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 10—figure supplement 6. Wavelet power spectra for national time series of neuropsychiatric disorder death rates for 1980–2016, by age group for females.

Figure 10—figure supplement 6.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.
Open in a new tab

Figure 3. Wavelet power spectra for national time series of cancer death rates for 1980–2016, by age group for males.

Figure 3.

Open in a new tab

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 4. Wavelet power spectra for national time series of cancer death rates for 1980–2016, by age group for females.

Figure 4.

Open in a new tab

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 5. Wavelet power spectra for national time series of cardiorespiratory disease death rates for 1980–2016, by age group for males.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 5.

Figure 5—figure supplement 1. Wavelet power spectra for national time series of cardiovascular disease death rates for 1980–2016, by age group for males.

Figure 5—figure supplement 1.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 5—figure supplement 2. Wavelet power spectra for national time series of chronic respiratory disease death rates for 1980–2016, by age group for males.

Figure 5—figure supplement 2.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 5—figure supplement 3. Wavelet power spectra for national time series of respiratory infection death rates for 1980–2016, by age group for males.

Figure 5—figure supplement 3.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.
Open in a new tab

Figure 6. Wavelet power spectra for national time series of cardiorespiratory disease death rates for 1980–2016, by age group for females.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 6.

Figure 6—figure supplement 1. Wavelet power spectra for national time series of cardiovascular disease death rates for 1980–2016, by age group for females.

Figure 6—figure supplement 1.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 6—figure supplement 2. Wavelet power spectra for national time series of chronic respiratory disease death rates for 1980–2016, by age group for females.

Figure 6—figure supplement 2.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 6—figure supplement 3. Wavelet power spectra for national time series of respiratory infection death rates for 1980–2016, by age group for females.

Figure 6—figure supplement 3.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.
Open in a new tab

Figure 7. Wavelet power spectra for national time series of injury death rates for 1980–2016, by age group for males.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 7.

Figure 7—figure supplement 1. Wavelet power spectra for national time series of intentional injury death rates for 1980–2016, by age group for males.

Figure 7—figure supplement 1.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 7—figure supplement 2. Wavelet power spectra for national time series of unintentional injury death rates for 1980–2016, by age group for males.

Figure 7—figure supplement 2.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.
Open in a new tab

Figure 8. Wavelet power spectra for national time series of injury death rates for 1980–2016, by age group for females.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 8.

Figure 8—figure supplement 1. Wavelet power spectra for national time series of intentional injury death rates for 1980–2016, by age group for females.

Figure 8—figure supplement 1.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 8—figure supplement 2. Wavelet power spectra for national time series of unintentional injury death rates for 1980–2016, by age group for females.

Figure 8—figure supplement 2.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.
Open in a new tab

Figure 9. Wavelet power spectra for national time series of death rates from causes other than cancers, cardiorespiratory diseases and injuries for 1980–2016, by age group for males.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 9.

Figure 9—figure supplement 1. Wavelet power spectra for national time series of substance use disorder death rates for 1980–2016, by age group for males.

Figure 9—figure supplement 1.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 9—figure supplement 2. Wavelet power spectra for national time series of perinatal condition death rates for 1980–2016, by age group for males.

Figure 9—figure supplement 2.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 9—figure supplement 3. Wavelet power spectra for national time series of endocrine disorder death rates for 1980–2016, by age group for males.

Figure 9—figure supplement 3.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 9—figure supplement 4. Wavelet power spectra for national time series of genitourinary disease death rates for 1980–2016, by age group for males.

Figure 9—figure supplement 4.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.

Figure 9—figure supplement 5. Wavelet power spectra for national time series of neuropsychiatric disorder death rates for 1980–2016, by age group for males.

Figure 9—figure supplement 5.

Wavelet power values increase from blue to red. The shaded regions at the left and right edge of each box indicate the cone of influence, where spectral analysis is less robust. P-values for the presence of 12 month seasonality are to the right of each figure at the 12 month line.
Open in a new tab

Centre of gravity analysis showed that death rates in men aged ≥45 years and women aged ≥35 years peaked in December, January or February and were lowest in June to August, for all-cause mortality as well as for all non-injury and non-maternal causes of death (Figure 11 and respective figure supplements). Deaths from cardiorespiratory diseases, including cardiovascular diseases, chronic respiratory diseases and respiratory infections, were also consistently highest in January and February and lowest in July and August across all ages, except for chronic respiratory diseases in ages 5–24 years where there are few deaths from this cause leading to unstable estimates (p-values for seasonality from wavelet analysis ranged from 0.35 to 0.48 for these ages). A similar temporal pattern was seen for all-cause and non-injury mortality in children younger than five years of age, whose all-cause death rate was highest in February and lowest in August. In contrast, among males aged 5–34 years, all-cause mortality peaked in June or July, as did deaths from injuries, which generally had a summer peak in males and females below 45 years of age.

Figure 11. Mean timing of maximum and minimum all-cause and cause-specific mortality at the national level, by sex and age group for 1980–2016.

Red arrows indicate the month of maximum mortality, and green arrows that of minimum mortality. The size of the arrow is inversely proportional to its respective variance.

Figure 11.

Figure 11—figure supplement 1. Mean timing of maximum and minimum mortality for specific cardiorespiratory diseases at the national level, by sex and age group for 1980–2016.

Figure 11—figure supplement 1.

Red arrows indicate the month of maximum mortality, and green arrows that of minimum mortality. The size of the arrow is inversely proportional to its respective variance.

Figure 11—figure supplement 2. Mean timing of maximum and minimum mortality for specific injuries at the national level, by sex and age group for 1980–2016.

Figure 11—figure supplement 2.

Red arrows indicate the month of maximum mortality, and green arrows that of minimum mortality. The size of the arrow is inversely proportional to its respective variance.

Figure 11—figure supplement 3. Mean timing of maximum and minimum mortality for the cluster of causes other than cancers, cardiorespiratory diseases and injuries at the national level, by sex and age group for 1980–2016.

Figure 11—figure supplement 3.

Red arrows indicate the month of maximum mortality, and green arrows that of minimum mortality. The size of the arrow is inversely proportional to its respective variance.
Open in a new tab

From 1980 to 2016, the proportional (percent) difference in all-cause death rates between peak and minimum months declined little for people older than 45 years of age (by less than eight percentage points with p-values for declining trend >0.1) (Figure 12). In contrast, the difference between peak (summer) and minimum (winter) death rates declined in younger ages, by over 25 percentage points in males aged 5–14 years and 15–24 years (p-values<0.01), largely driven in the declining difference between summer and winter injury deaths. Under five years of age, percent seasonal difference in all-cause death rates declined by 13 percentage points (p-value<0.01) for boys but only five percentage points (p-value=0.12) for girls. These declines in seasonality of child deaths were a net effect of declining winter-summer difference in cardiorespiratory diseases deaths and increasing summer-winter difference in injury deaths, itself driven by increasing difference in non-intentional injuries (Figure 12—figure supplement 1). Within specific cardiorespiratory diseases in under-five children, percent difference declined for cardiorespiratory diseases, cardiovascular diseases, and chronic respiratory diseases while increasing for respiratory infections.

Figure 12. National percent difference in death rates between the maximum and minimum mortality months for all-cause and cause-specific mortality in 2016 versus 1980, by sex and age group.

Figure 12.

Figure 12—figure supplement 1. National percent difference in death rates between the maximum and minimum mortality months for specific injuries in 2016 versus 1980, by sex and age group.

Figure 12—figure supplement 1.

Figure 12—figure supplement 2. National percent difference in death rates between the maximum and minimum mortality months for specific cardiorespiratory diseases in 2016 versus 1980, by sex and age group.

Figure 12—figure supplement 2.

Figure 12—figure supplement 3. National percent difference in death rates between the maximum and minimum mortality months for the cluster of causes other than cancers, cardiorespiratory diseases and injuries in 2016 versus 1980, by sex and age group.

Figure 12—figure supplement 3.

Open in a new tab

The subnational centre of gravity analysis showed that all-cause mortality peaks and minima in different climate regions are consistent with the national ones (Figures 13–16), indicating that seasonality is largely independent of geography. The relative homogeneity of the timing of maximum and minimum mortality contrasts with the large variation in seasonal temperatures among climate regions. For example, in men and women aged 65–74 years, all-cause mortality peaked in February in the Northeast and Southeast, even though the average temperatures for those regions were different by over 13 degrees Celsius (9.3 in the Southeast compared with −3.8 in the Northeast). Furthermore, above 45 years of age, there was little inter-region variation in the percent seasonal difference in all-cause mortality, despite the large variation in temperature difference between the peak and minimum months (Figure 17).

Figure 13. Mean timing of maximum all-cause mortality for 1980–2016, by climate region and age group for males.

Figure 13.

Open in a new tab

Average temperatures (in degrees Celsius) are included in white for the corresponding month of maximum and minimum mortality for each climate region.

Figure 16. Mean timing of minimum all-cause mortality for 1980–2016, by climate region and age group for females.

Figure 16.

Open in a new tab

Average temperatures (in degrees Celsius) are included in white for the corresponding month of maximum and minimum mortality for each climate region.

Figure 17. The relationship between percent difference in all-cause death rates and temperature difference between months with maximum and minimum mortality across climate regions, by sex and age group in 2016.

Figure 17.

Open in a new tab

Figure 14. Mean timing of minimum all-cause mortality for 1980–2016, by climate region and age group for males.

Figure 14.

Open in a new tab

Average temperatures (in degrees Celsius) are included in white for the corresponding month of maximum and minimum mortality for each climate region.

Figure 15. Mean timing of maximum all-cause mortality for 1980–2016, by climate region and age group for females.

Figure 15.

Open in a new tab

Average temperatures (in degrees Celsius) are included in white for the corresponding month of maximum and minimum mortality for each climate region.

Strengths and limitations

The strengths of our study are its innovative methods of characterizing seasonality of mortality dynamically over space and time, by age group and cause of death; using wavelet and centre of gravity analyses; using ERA-Interim data output to compare the association between seasonality of death rates and regional temperature. A limitation of our study is that we did not investigate seasonality of mortality by socioeconomic characteristics which may help with understanding its determinants and planning responses.

Discussion

We used wavelet and centre of gravity analyses, which allowed systematically identifying and characterizing seasonality of total and cause-specific mortality in the USA, and examining how seasonality has changed over time. We identified distinct seasonal patterns in relation to age and sex, including higher all-cause summer mortality in young men (Feinstein, 2002; Rau et al., 2018). Importantly, we also showed that all-cause and cause-specific mortality seasonality is largely similar in terms of both timing and magnitude across diverse climatic regions with substantially different summer and winter temperatures. Insights of this kind would not have been possible analysing data averaged over time or nationally, or fixed to pre-specified frequencies.

Prior studies have noted seasonality of mortality for all-cause mortality and for specific causes of death in the USA (Feinstein, 2002; Kalkstein, 2013; Rau, 2004; Rau et al., 2018; Rosenwaike, 1966; Seretakis et al., 1997). Few of these studies have done consistent national and subnational analyses, and none has done so over time, for a comprehensive set of age groups and causes of death, and in relation to regional temperature differences. Our results on strong seasonality of cardiorespiratory diseases deaths and weak seasonality of cancer deaths, restricted to older ages, are broadly consistent with these studies (Feinstein, 2002; Rau et al., 2018; Rosenwaike, 1966; Seretakis et al., 1997), which had limited analysis on how seasonality changes over time and geography (Feinstein, 2002; Rau et al., 2018; Rosenwaike, 1966). Similarly, our results on seasonality of injury deaths are supported by a few prior studies (Feinstein, 2002; Rau et al., 2018; Rosenwaike, 1966), but our subnational analysis over three decades revealed variations in when injury deaths peaked and in how seasonal differences in these deaths have changed over time in relation to age group which had not been reported before.

A study of 36 cities in the USA, aggregated across age groups and over time, also found that excess mortality was not associated with seasonal temperature range (Kinney et al., 2015). In contrast, a European study found that the difference between winter and summer mortality was lower in colder Nordic countries than in warmer southern European nations (Healy, 2003; McKee, 1989) (the study’s measure of temperature was mean annual temperature which differed from the temperature difference between maximum and minimum mortality used in our analysis although the two measures are correlated). The absence of variation in the magnitude of mortality seasonality indicates that different regions in the USA are similarly adapted to temperature seasonality, whereas Nordic countries may have better environmental (e.g. housing insulation and heating) and health system measures to counter the effects of cold winters than those in southern Europe. If the observed absence of association between the magnitude of mortality seasonality and seasonal temperature difference across the climate regions also persists over time, the changes in temperature as a result of global climate change are unlikely to affect the winter-summer mortality difference.

The cause-specific analysis showed that the substantial decline in seasonal mortality differences in adolescents and young adults was related to the diminishing seasonality of (unintentional) injuries, especially from road traffic crashes, which are more likely to occur in the summer months (Liu et al., 2005) and are more common in men. The weakening of seasonality in boys under five years of age was related to two phenomena: first, the seasonality of death from cardiorespiratory diseases declined, and second, the proportion of deaths from perinatal conditions, which exhibit limited seasonality (Figure 9—figure supplement 2 and Figure 10—figure supplement 3), increased (MacDorman and Gregory, 2015).

In contrast to young and middle ages, mortality in older ages, where death rates are highest, maintained persistent seasonality over a period of three decades (we note that although the percent seasonal difference in mortality has remained largely unchanged in these ages, the absolute difference in death rates between the peak and minimum months has declined because total mortality has a declining long-term trend). This finding demonstrates the need for environmental and health service interventions targeted towards this group irrespective of geography and local climate. Examples of such interventions include enhancing the availability of both environmental and medical protective factors, such as better insulation of homes, winter heating provision and flu vaccinations, for the vulnerable older population (Katiyo et al., 2017). Social interventions, including regular visits to the isolated elderly during peak mortality periods to ensure that they are optimally prepared for adverse conditions, and responsive and high-quality emergency care, are also important to protect this vulnerable group (Healy, 2003; Lerchl, 1998; Katiyo et al., 2017). Emergent new technologies, such as always-connected hands-free communications devices with the outside world, in-house cameras, and personal sensors also provide an opportunity to enhance care for the older, more vulnerable groups in the population, especially in winter when the elderly have fewer social interactions (Morris, 2013). Such interventions are important today, and will remain so as the population ages and climate change increases the within- and between-season weather variability.

Materials and methods

Data

We used data on all 85,854,176 deaths in the USA from 1980 to 2016 from the National Center for Health Statistics (NCHS). Age, sex, state of residence, month of death, and underlying cause of death were available for each record. The underlying cause of death was coded according to the international classification of diseases (ICD) system (9th revision of ICD from 1980 to 1998 and 10th revision of ICD thereafter). Yearly population counts were available from NCHS for 1990 to 2016 and from the US Census Bureau prior to 1990 (Ingram et al., 2003). We calculated monthly population counts through linear interpolation, assigning each yearly count to July.

We also subdivided the national data geographically into nine climate regions used by the National Oceanic and Atmospheric Administration (Figure 18 and Table 2) (Karl and Koss, 1984). On average, the Southeast and South are the hottest climate regions with average annual temperatures of 18.4°C and 18°C respectively; the South also possesses the highest average maximum monthly temperature (27.9°C in July). The lowest variation in temperature throughout the year is that of the Southeast (an average range of 17.5°C). The three coldest climate regions are West North Central, East North Central and the Northwest (7.6°C, 8.0°C, 8.2°C respectively). Mirroring the characteristics of the hottest climate regions, the largest variation in temperature throughout the year is that of the coldest region, West North Central (an average range of 30.5°C), which also has the lowest average minimum monthly temperature (−6.5°C in January). The other climate regions, Northeast, Southwest, and Central, possess similar average temperatures (10°C to 14°C) and variation within the year of (23°C to 26°C), with the Northeast being the most populous region in the United States (with 19.8% total population in 2016).

Figure 18. Climate regions of the USA.

Figure 18.

Open in a new tab

Table 2. Characteristics of climate regions of the USA.

Climate region Constituent states Population (2016) Mean annual temperature (1980–2016) (°C)
Central Illinois, Indiana, Kentucky, Missouri, Ohio, Tennessee, West Virginia 50,191,326 11.6
East North Central Iowa, Michigan, Minnesota, Wisconsin 24,418,738 8
Northeast Connecticut, Delaware, Maine, Maryland, Massachusetts, New Hampshire, New Jersey, New York, Pennsylvania, Rhode Island, Vermont 64,046,741 10.6
Northwest Alaska, Idaho, Oregon, Washington 13,811,810 8.2
South Arkansas, Kansas, Louisiana, Mississippi, Oklahoma, Texas 45,388,414 18
Southeast Alabama, Florida, Georgia, North Carolina, South Carolina, Virginia 59,356,072 18.4
Southwest Arizona, Colorado, New Mexico, Utah 17,613,981 13.6
West California, Hawaii, Nevada 43,708,574 16.6
West North Central Montana, Nebraska, North Dakota, South Dakota, Wyoming 5,168,753 7.6
Open in a new tab

Data were divided by sex and age in the following 10 age groups: 0–4, 5–14, 15-24, 25–34, 35–44, 45–54, 55–64, 65–74, 75–84, 85+ years. We calculated monthly death rates for each age and sex group, both nationally and for sub-national climate regions. Death rate calculations accounted for varying length of months, by multiplying each month’s death count by a factor that would make it equivalent to a 31 day month.

For analysis of seasonality by cause of death, we mapped each ICD-9 and ICD-10 codes to four main disease categories (Table 1) and to a number of subcategories which are presented in the Supplementary Note. Cardiorespiratory diseases and cancers accounted for 56.4% and 21.2% of all deaths in the USA, respectively, in 1980, and 40.3% and 22.4%, respectively, in 2016. Deaths from cardiorespiratory diseases have been associated with cold and warm temperatures (Basu, 2009; Basu and Samet, 2002; Bennett et al., 2014; Braga et al., 2002; Gasparrini et al., 2015). Injuries, which accounted for 8% of all deaths in the USA in 1980 and 7.3% in 2016, may have seasonality that is distinct from so-called natural causes. We did not further divide other causes because the number of deaths could become too small to allow stable estimates when divided by age group, sex and climate region.

We obtained data on temperature from ERA-Interim, which combines predictions from a physical model with ground-based and satellite measurements (Dee et al., 2011). We used gridded four-times-daily estimates at a resolution of 80 km to generate monthly population-weighted temperature by climate region throughout the analysis period.

Statistical methods

We used wavelet analysis to investigate seasonality for each age-sex group. Wavelet analysis uncovers the presence, and frequency, of repeated maxima and minima in each age-sex-specific death rate time series (Hubbard, 1998; Torrence and Compo, 1998). In brief, a Morlet wavelet, described in detail elsewhere (Cazelles et al., 2008), is equivalent to using a moving window on the death rate time series and analysing periodicity in each window using a short-form Fourier transform, hence generating a dynamic spectral analysis, which allows measuring dynamic seasonal patterns, in which the periodicity of death rates may disappear, emerge, or change over time. In addition to coefficients that measure the frequency of periodicity, wavelet analysis estimates the probability of whether the data are different from the null situation of random fluctuations that can be represented with white (an independent random process) or red (autoregressive of order one process) noise. For each age-sex group, we calculated the p-values of the presence of 12 month seasonality for the comparison of wavelet power spectra of the entire study period (1980–2016) with 100 simulations against a white noise spectrum, which represents random fluctuations. We used the R package WaveletComp (version 1.0) for the wavelet analysis. Before analysis, we de-trended death rates using a polynomial regression, and rescaled each death rate time series so as to range between 1 and −1.

To identify the months of maximum and minimum death rates, we calculated the centre of gravity and the negative centre of gravity of monthly death rates. Centre of gravity was calculated as a weighted average of months of deaths, with each month weighted by its death rate; negative centre of gravity was also calculated as a weighted average of months of deaths, but with each month was weighted by the difference between its death rate and the year’s maximum death rate. In taking the weighted average, we allowed December (month 12) to neighbour January (month 1), representing each month by an angle subtended from 12 equally-spaced points around a unit circle. Using a technique called circular statistics, a mean (θ-)of the angles (θ1,θ2,θ3…,θn,) representing the deaths (with n the total number of deaths in an age-sex group for a particular cause of death) is found using the relation below:

θ-=arg∑j=1nexp⁡(iθj),

where arg denotes the complex number argument and θj denotes the month of death in angular form for a particular death j. The outcome of this calculation is then converted back into a month value (Fisher, 1995). Along with each circular mean, a 95% confidence interval (CI) was calculated by using 1000 bootstrap samples. The R package CircStats (version 0.2.4) was used for this analysis.

For each age-sex group and cause of death, and for each year, we calculated the percent difference in death rates between the maximum and minimum mortality months. We fitted a linear regression to the time series of seasonal differences from 1980 to 2016, and used the fitted trend line to estimate how much the percentage difference in death rates between the maximum and minimum mortality months had changed from 1980 to 2016. We weighted seasonal difference by the inverse of the square of its standard error, which was calculated using a Poisson model to take population size of each age-sex group through time into account. This method gives us a p-value for the change in seasonal difference per year, which we used to calculate the seasonal difference at the start (1980) and end (2016) of the period of study. Our method of analysing seasonal differences avoids assuming that any specific month or group of months represent highest and lowest number of deaths for a particular cause of death, which is the approach taken by the traditional measure of Excess Winter Deaths. It also allows the maximum and minimum mortality months to vary by age group, sex and cause of death.

Acknowledgments

Robbie Parks is supported by a Wellcome Trust ISSF Studentship. Work on the US mortality data is supported by a grant from US Environmental Protection Agency. The views expressed in this document are solely those of the authors and do not necessarily reflect those of the Agency.

Funding Statement

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Contributor Information

Majid Ezzati, Email: majid.ezzati@imperial.ac.uk.

Eduardo Franco, McGill University, Canada.

Prabhat Jha, Saint Michael's Hospital, Canada.

Funding Information

This paper was supported by the following grants:

  • Wellcome Trust 205208/Z/16/Z to Majid Ezzati.

  • U.S. Environmental Protection Agency RD-83587301 to Majid Ezzati.

Additional information

Competing interests

No competing interests declared.

Reports a charitable grant from the AstraZeneca Young Health Programme, and personal fees from Prudential, Scor, and Third Bridge, all outside the submitted work.

Author contributions

Conceptualization, Data curation, Software, Formal analysis, Investigation, Visualization, Methodology, Writing—original draft.

Conceptualization, Visualization, Methodology, Writing—review and editing.

Data curation, Methodology, Writing—review and editing.

Conceptualization, Methodology, Writing—review and editing.

Conceptualization, Resources, Data curation, Supervision, Funding acquisition, Visualization, Methodology, Writing—original draft, Project administration.

Additional files

Transparent reporting form
elife-35500-transrepform.docx (246.2KB, docx)
DOI: 10.7554/eLife.35500.050

Data availability

The ERA-Interim temperature data are available at https://www.ecmwf.int/en/forecasts/datasets/reanalysis-datasets/era-interim. The US Census Populations With Bridged Race Categories data for 1990-2016 are available at https://wonder.cdc.gov/bridged-race-population.html. Pre 1990, the County Intercensal Tables are available at https://www.census.gov/data/tables/time-series/demo/popest/1980s-county.html. The Vital statistics data are available at https://www.cdc.gov/nchs/nvss/dvs_data_release.htm through a request to NAPHSIS (https://www.naphsis.org/).

The following previously published datasets were used:

National Center for Health Statistics. 2017. Vital statistics (1980-2016) National Center for Health Statistics. dvs_data_release

European Centre for Medium-Range Weather Forecasts. 2016. ERA-Interim temperature data (1979-2016) European Centre for Medium-Range Weather Forecasts. reanalysis-datasets/era-interim

CDC Wonder. 2016. US Census Populations With Bridged Race Categories (1990-2016) CDC WONDER. bridged-race-population

United States Census Bureau. 2016. County Intercensal Tables (1980-1989) United States Census Bureau. 1980s-county

References

  1. Basu R, Samet JM. Relation between elevated ambient temperature and mortality: a review of the epidemiologic evidence. Epidemiologic Reviews. 2002;24:190–202. doi: 10.1093/epirev/mxf007. [DOI] [PubMed] [Google Scholar]
  2. Basu R. High ambient temperature and mortality: a review of epidemiologic studies from 2001 to 2008. Environmental Health. 2009;8:40. doi: 10.1186/1476-069X-8-40. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Bennett JE, Blangiardo M, Fecht D, Elliott P, Ezzati M. Vulnerability to the mortality effects of warm temperature in the districts of England and Wales. Nature Climate Change. 2014;4:269–273. doi: 10.1038/nclimate2123. [DOI] [Google Scholar]
  4. Bobb JF, Peng RD, Bell ML, Dominici F. Heat-related mortality and adaptation to heat in the United States. Environmental Health Perspectives. 2014;122:811–816. doi: 10.1289/ehp.1307392. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Braga AL, Zanobetti A, Schwartz J. The time course of weather-related deaths. Epidemiology. 2001;12:662–667. doi: 10.1097/00001648-200111000-00014. [DOI] [PubMed] [Google Scholar]
  6. Braga AL, Zanobetti A, Schwartz J. The effect of weather on respiratory and cardiovascular deaths in 12 U.S. cities. Environmental Health Perspectives. 2002;110:859–863. doi: 10.1289/ehp.02110859. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Campbell A. Excess winter mortality in England and Wales: 2016 to 2017 (provisional) and 2015 to 2016 (final) Statistical Bulletin, Office for National Statistics 2017 [Google Scholar]
  8. Carson C, Hajat S, Armstrong B, Wilkinson P. Declining vulnerability to temperature-related mortality in London over the 20th century. American Journal of Epidemiology. 2006;164:77–84. doi: 10.1093/aje/kwj147. [DOI] [PubMed] [Google Scholar]
  9. Cazelles B, Chavez M, Berteaux D, Ménard F, Vik JO, Jenouvrier S, Stenseth NC. Wavelet analysis of ecological time series. Oecologia. 2008;156:287–304. doi: 10.1007/s00442-008-0993-2. [DOI] [PubMed] [Google Scholar]
  10. Davis RE, Knappenberger PC, Michaels PJ, Novicoff WM. Seasonality of climate-human mortality relationships in US cities and impacts of climate change. Climate Research. 2004;26:61–76. doi: 10.3354/cr026061. [DOI] [Google Scholar]
  11. Dee DP, Uppala SM, Simmons AJ, Berrisford P, Poli P, Kobayashi S, Andrae U, Balmaseda MA, Balsamo G, Bauer P, Bechtold P, Beljaars ACM, van de Berg L, Bidlot J, Bormann N, Delsol C, Dragani R, Fuentes M, Geer AJ, Haimberger L, Healy SB, Hersbach H, Hólm EV, Isaksen L, Kållberg P, Köhler M, Matricardi M, McNally AP, Monge-Sanz BM, Morcrette J-J, Park B-K, Peubey C, de Rosnay P, Tavolato C, Thépaut J-N, Vitart F. The ERA-Interim reanalysis: configuration and performance of the data assimilation system. Quarterly Journal of the Royal Meteorological Society. 2011;137:553–597. doi: 10.1002/qj.828. [DOI] [Google Scholar]
  12. Feinstein CA. Seasonality of deaths in the U.S. by age and cause. Demographic Research. 2002;6:471–488. doi: 10.4054/DemRes.2002.6.17. [DOI] [Google Scholar]
  13. Fisher N. Statistical Analysis of Circular Data. Cambridge University Press; 1995. [Google Scholar]
  14. Fowler T, Southgate RJ, Waite T, Harrell R, Kovats S, Bone A, Doyle Y, Murray V. Excess winter deaths in Europe: a multi-country descriptive analysis. European Journal of Public Health. 2015;25:339–345. doi: 10.1093/eurpub/cku073. [DOI] [PubMed] [Google Scholar]
  15. Gasparrini A, Guo Y, Hashizume M, Lavigne E, Zanobetti A, Schwartz J, Tobias A, Tong S, Rocklöv J, Forsberg B, Leone M, De Sario M, Bell ML, Guo YL, Wu CF, Kan H, Yi SM, de Sousa Zanotti Stagliorio Coelho M, Saldiva PH, Honda Y, Kim H, Armstrong B. Mortality risk attributable to high and low ambient temperature: a multicountry observational study. The Lancet. 2015;386:369–375. doi: 10.1016/S0140-6736(14)62114-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Grenfell BT, Bjørnstad ON, Kappey J. Travelling waves and spatial hierarchies in measles epidemics. Nature. 2001;414:716–723. doi: 10.1038/414716a. [DOI] [PubMed] [Google Scholar]
  17. Healy JD. Excess winter mortality in Europe: a cross country analysis identifying key risk factors. Journal of Epidemiology & Community Health. 2003;57:784–789. doi: 10.1136/jech.57.10.784. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Hubbard BB. The World According to Wavelets: The Story of a Mathematical Technique in the Making. A K Peters/CRC Press; 1998. [Google Scholar]
  19. Ingram DD, Parker JD, Schenker N, Weed JA, Hamilton B, Arias E, Madans JH. United states census 2000 population with bridged race categories. Vital and Health Statistics. Series 2, Data Evaluation and Methods Research. 2003:1–55. [PubMed] [Google Scholar]
  20. Kalkstein AJ. Regional similarities in seasonal mortality across the United States: an examination of 28 metropolitan statistical areas. PLoS One. 2013;8:e63971. doi: 10.1371/journal.pone.0063971. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Karl TR, Koss WJ. Regional and National Monthly, Seasonal, and Annual Temperature Weighted by Area, 1895-1983. Asheville, United States: National Climatic Data Center; 1984. [Google Scholar]
  22. Katiyo S, Dorey S, Bone A. The Cold Weather Plan for England. Public Health England; 2017. [Google Scholar]
  23. Kinney PL, Schwartz J, Pascal M, Petkova E, Tertre AL, Medina S, Vautard R. Winter season mortality: will climate warming bring benefits? Environmental Research Letters. 2015;10:064016. doi: 10.1088/1748-9326/10/6/064016. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Langford IH, Bentham G. The potential effects of climate change on winter mortality in England and Wales. International Journal of Biometeorology. 1995;38:141–147. doi: 10.1007/BF01208491. [DOI] [PubMed] [Google Scholar]
  25. Lerchl A. Changes in the seasonality of mortality in Germany from 1946 to 1995: the role of temperature. International Journal of Biometeorology. 1998;42:84–88. doi: 10.1007/s004840050089. [DOI] [PubMed] [Google Scholar]
  26. Liu C, Chen C, Utter D. Trend and pattern analysis of highway crash fatality by month and day. National Highway Traffic Safety Administration Technical Report 2005 [Google Scholar]
  27. MacDorman MF, Gregory EC. Fetal and perinatal mortality: nited States, 2013. National Vital Statistics Reports : From the Centers for Disease Control and Prevention, National Center for Health Statistics, National Vital Statistics System. 2015;64:1–24. [PubMed] [Google Scholar]
  28. Martens WJ. Climate change, thermal stress and mortality changes. Social Science & Medicine. 1998;46:331–344. doi: 10.1016/S0277-9536(97)00162-7. [DOI] [PubMed] [Google Scholar]
  29. McKee CM. Deaths in winter: can Britain learn from Europe? European Journal of Epidemiology. 1989;5:178–182. doi: 10.1007/BF00156826. [DOI] [PubMed] [Google Scholar]
  30. Medina-Ramón M, Schwartz J. Temperature, temperature extremes, and mortality: a study of acclimatisation and effect modification in 50 US cities. Occupational and Environmental Medicine. 2007;64:827–833. doi: 10.1136/oem.2007.033175. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Morris ME. Smart-Home technologies to assist older people to live well at home. Journal of Aging Science. 2013;1:101. doi: 10.4172/2329-8847.1000101. [DOI] [Google Scholar]
  32. Moy CM, Seltzer GO, Rodbell DT, Anderson DM. Variability of El Niño/Southern Oscillation activity at millennial timescales during the Holocene epoch. Nature. 2002;420:162–165. doi: 10.1038/nature01194. [DOI] [PubMed] [Google Scholar]
  33. Rau R. Wirtschafts- Und Sozialwissenschaftlichen Fakultät. Vol. 361. New York City, United States: Springer-Verlag Berlin Heidelberg; 2004. Seasonality in Human Mortality. A Demographic Approach. [Google Scholar]
  34. Rau R, Bohk-Ewald C, Muszyńska MM, Vaupel JW. Visualizing Mortality Dynamics in the Lexis Diagram. Springer; 2018. [DOI] [PubMed] [Google Scholar]
  35. Rosenwaike I. Seasonal variation of deaths in the United States, 1951–1960. Journal of the American Statistical Association. 1966;61:706–719. [Google Scholar]
  36. Seretakis D, Lagiou P, Lipworth L, Signorello LB, Rothman KJ, Trichopoulos D. Changing seasonality of mortality from coronary heart disease. JAMA: The Journal of the American Medical Association. 1997;278:1012. doi: 10.1001/jama.1997.03550120072036. [DOI] [PubMed] [Google Scholar]
  37. Sheridan SC, Kalkstein AJ, Kalkstein LS. Trends in heat-related mortality in the United States, 1975–2004. Natural Hazards. 2009;50:145–160. doi: 10.1007/s11069-008-9327-2. [DOI] [Google Scholar]
  38. Torrence C, Compo GP. A practical guide to wavelet analysis. Bulletin of the American Meteorological Society. 1998;79:61–78. doi: 10.1175/1520-0477(1998)079&#x0003c;0061:APGTWA&#x0003e;2.0.CO;2. [DOI] [Google Scholar]

Decision letter

Editor: Eduardo Franco1
Reviewed by: Tom Fowler2, Patrick Brown3

In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.

Thank you for submitting your article "National and regional seasonal dynamics of all-cause and cause-specific mortality in the USA from 1980 to 2013" for consideration by eLife. Your article has been reviewed by three peer reviewers, and the evaluation has been overseen by Eduardo Franco as a Reviewing Editor and Prabhat Jha as the Senior Editor. The Reviewing Editor has drafted this decision to help you prepare a revised submission. As is common in eLife, decision narratives are an amalgamation of the essential points of the reviewers' critiques after eliminating redundancies in comments and suggestions.

Summary:

This paper examines the seasonality in age- and disease-specific mortality in the US using geocoded data analyzed via wavelet statistical modelling and Poisson regression techniques. The key results indicate mortality increases in adults over 45 years old in winter from cardiorespiratory causes and injuries. Injuries, attributable mainly to road traffic crashes were seen to rise in the summer months among young men. No variation in seasonality was found by climate region, indicating no apparent disparities in adaptation by geographic location. This study provides insights as to whether climate changes could have an impact on the mortality patterns observed in the US as a model. The findings could assist in planning surveillance mechanisms and in projecting workforce requirements in primary and emergency healthcare.

Essential revisions:

1) By choosing only 4 disease groups the authors seem to have missed a key opportunity to investigate seasonal patterns in finer subgroups despite possessing a large sample size (n = 77,771,264), Furthermore, the rationale for choosing the subgroups presented is unclear. For example, the authors chose to group cardiorespiratory diseases into one large outcome when cardiovascular and respiratory outcomes have differing mechanisms in their association with heat or cold exposure. Within respiratory diseases itself, it makes sense to separate acute and chronic causes because the mechanisms vary so starkly and are related to seasonal patterns, i.e. the incidence of pneumonia and acute respiratory outcomes are likely heightened in the winter and this observation may differ to how/why the incidence of chronic respiratory deaths vary seasonally. Furthermore, the mechanisms for how heat is associated with elevated respiratory deaths is understudied and knowing more about seasonal patterns here would be a useful addition to the literature.

I believe it's important to investigate seasonal patterns of deaths from infectious disease, maternal and neonatal causes, endocrine disorders, genitourinary conditions and neuropsychiatric conditions (which are all associated with temperature) to add valuable insight into understudied but important disease groups, but these have been excluded from the present analysis. Considering the tragic opioid epidemic in USA, it would be a valuable contribution to the literature to understand whether seasonal patterns (at least in more recent years) are observed with deaths related to substance use disorders.

2) The authors pointed out in the Introduction that global warming may impact on excess cold weather death rates. Presumably a major part of the rationale for testing the differences in regions over time was to help inform our understanding of what the impact may be. However no conclusions were reached with regard to the implications of the findings. Clearly any such conclusions would need to be appropriately and heavily caveated but as this is a major reason for the analysis. Please expand the Discussion to include these points.

3) Methodology:

A) Wavelet power spectra are not always easy to interpret, and the uncertainty in estimated wavelet coefficients is difficult to quantify. The wavelet analysis makes for an interesting exploratory tool, showing that for the most part cycles have a duration of 12 months and are reasonably stable over time. It does not seem possible to draw firm conclusions about the research hypothesis using wavelets, however. All cause male mortality for 15-24 year olds appears to be less cyclical in recent years, although how to quantify this effect and assign a statistical significance to it is not apparent from the power spectrum shown.

The second analysis is more appropriate, although the details of this analysis are sparse. It appears that the model used is something like the following, where Yit is the count of deaths in age group i at time t.

--------------------- -------------------------------------

Yit∼ Poisson(Nit ⋅ λit)log(λit)= μi + βit + Mit ⋅ [α1i ⋅ cos(2πt/12)

+α2i ⋅ sin(2πt/12)+α3i ⋅ cos(2πt/6)

+α4i ⋅ sin(2πt/6)]

Mit= ρi + γit

--------------------- -------------------------------------

A model of this could answer the following questions, with p-values produced from a likelihood ratio test.

- Is γi negative? If so the seasonal effect is becoming less severe for age group i.

- Does it look like all age groups have the same trend, with γi = γ₀ for all i?

- Do all age groups have the same cycle, with αpi = αp0?

- Does γ vary by region (negative in cold-climate regions?)

- Do the cycles, given by the α, vary by region or is a nation-wide cycle sufficient?

The above model is easier to interpret than the wavelet analysis. The quantity 1 − exp(120γ) is the change per decade in the seasonality effect, which could be reported with a 95% confidence interval.

B) The approach is different to a number of those referenced (particularly the standard measure of excess winter deaths). It would be helpful to have some comment on this, particularly the generalisability to other studies. The focus has been on peak months in this analysis, but the standard assessment of excess winter deaths is to compare December to March months to the rest of the year.

C) Appropriate rationale was given to the regional splits used in the analysis. However no description was given of the characteristics of these regional areas other than subsequent information on temperatures. As pointed out there could be a number of factors that are important. I would expect some reference to their differing characteristics.

4) Interpretation:

A) No table is presented with descriptive statistics, i.e. on the sample size of cases that fall into each disease category.

B) I may be misreading the paper (in which case it would be helpful to clarify so others do not make the same mistake), however the analysis in Figure 6 compares the difference in temperature experienced in those regions between the warmest and coldest months versus the% seasonal difference in death. In the Discussion this in contrasted with findings from Europe where countries with a more temperate winter have, paradoxically, higher rates of excess winter mortality. However this is not an appropriate direct comparison, regional extremes in variation were not looked at when comparisons between countries are made. This does not invalidate the comparison but it would be helpful to explicitly summarise if there are any differences between regions and examine what those are.

C) One option for this paper would be to be admittedly exploratory, avoiding the use of the word 'significant' and simplifying the analysis. Simple monthly averages and testing for the months having the same mean could replace the wavelet analysis. A second option would be to focus on a specific research hypothesis, explain carefully how the model estimates relate to this research hypothesis, and adjust the p-values for multiple testing.

eLife. 2018 Oct 30;7:e35500. doi: 10.7554/eLife.35500.063

Author response

Essential revisions:

1) By choosing only 4 disease groups the authors seem to have missed a key opportunity to investigate seasonal patterns in finer subgroups despite possessing a large sample size (n = 77,771,264), Furthermore, the rationale for choosing the subgroups presented is unclear. For example, the authors chose to group cardiorespiratory diseases into one large outcome when cardiovascular and respiratory outcomes have differing mechanisms in their association with heat or cold exposure. Within respiratory diseases itself, it makes sense to separate acute and chronic causes because the mechanisms vary so starkly and are related to seasonal patterns, i.e. the incidence of pneumonia and acute respiratory outcomes are likely heightened in the winter and this observation may differ to how/why the incidence of chronic respiratory deaths vary seasonally. Furthermore, the mechanisms for how heat is associated with elevated respiratory deaths is understudied and knowing more about seasonal patterns here would be a useful addition to the literature.

I believe it's important to investigate seasonal patterns of deaths from infectious disease, maternal and neonatal causes, endocrine disorders, genitourinary conditions and neuropsychiatric conditions (which are all associated with temperature) to add valuable insight into understudied but important disease groups, but these have been excluded from the present analysis. Considering the tragic opioid epidemic in USA, it would be a valuable contribution to the literature to understand whether seasonal patterns (at least in more recent years) are observed with deaths related to substance use disorders.

We had selected the four cause groups to go beyond all-cause mortality and still have parsimonious presentation and sufficient number of events by sex, age group, year, month and geography. In the specific case of cardiorespiratory diseases, they are commonly analysed together when studying the effects of temperature and air pollution (Basu, 2009; Basu, Dominici, and Samet, 2005; Basu and Samet, 2002; Bennett et al., 2014; Braga, Zanobetti, and Schwartz, 2002; Curriero et al., 2002; Dockery et al., 1993; Gasparrini et al., 2015; Hoek et al., 2013; Pope et al., 2002), possibly because of their shared aetiology (e.g., death from cor pulmonale arising from COPD or death from pneumonia in patients with acute vascular events) and because assignment of cause of death may use them interchangeably. We now present a third layer of disaggregation following the above suggestions (figure supplements for Figures 6-13; Table 1; Materials and methods).

2) The authors pointed out in the Introduction that global warming may impact on excess cold weather death rates. Presumably a major part of the rationale for testing the differences in regions over time was to help inform our understanding of what the impact may be. However no conclusions were reached with regard to the implications of the findings. Clearly any such conclusions would need to be appropriately and heavily caveated but as this is a major reason for the analysis. Please expand the Discussion to include these points.

To quantitatively analyse the potential implications of global climate change, one would need to formally incorporate temperature in the analysis which would involve a distinct analytical framework and presentation. As correctly pointed out in this comment, the current analysis can nonetheless provide qualitative insights into such effects, which we have now discussed with appropriate caveats (Discussion, second, third and fourth paragraphs).

3) Methodology:

A) Wavelet power spectra are not always easy to interpret, and the uncertainty in estimated wavelet coefficients is difficult to quantify. The wavelet analysis makes for an interesting exploratory tool, showing that for the most part cycles have a duration of 12 months and are reasonably stable over time. It does not seem possible to draw firm conclusions about the research hypothesis using wavelets, however. All cause male mortality for 15-24 year olds appears to be less cyclical in recent years, although how to quantify this effect and assign a statistical significance to it is not apparent from the power spectrum shown.

Wavelet analysis is a formal analytical framework that can provide quantitative conclusions (Hubbard, 1996; Torrence and Compo, 1998), and has been applied in different areas of health and the environment, for example measles epidemics and El Niño oscillations, to quantitatively characterise their seasonality including.(Grenfell, Bjørnstad, and Kappey, 2001; Moyet al., 2002) A key advantage of wavelet analysis is that it does not assume a stationary time series, and hence can easily capture a decline/increase in or appearance/disappearance of seasonality. The uncertainty of wavelet spectra are also well-characterised (Hubbard, 1996). We follow wavelet analysis with a centre of gravity analysis to identify the months of maximum and minimum mortality (Figure 12 and respective figure supplements; subsection “Statistical methods"), and with a quantitative analysis of (percentage) difference in death rates from 1980 to 2016 (Figure 13 and respective figure supplements; subsection “Statistical methods").

We note that even in standard statistical comparisons, p-values and statistical significance are not apparent from figures but are a part of the analysis. We have added p-values to each of the wavelet analyses in Figures 2-11 and the respective figure supplements (subsection “Statistical methods").

The second analysis is more appropriate, although the details of this analysis are sparse. It appears that the model used is something like the following, where Yit is the count of deaths in age group i at time t.

--------------------- -------------------------------------

Yit∼ Poisson(Nit ⋅ λit)log(λit)= μi + βit + Mit ⋅ [α1i ⋅ cos(2πt/12)

+α2i ⋅ sin(2πt/12)+α3i ⋅ cos(2πt/6)

+α4i ⋅ sin(2πt/6)]

Mit= ρi + γit

--------------------- -------------------------------------

A model of this could answer the following questions, with p-values produced from a likelihood ratio test.

- Is γi negative? If so the seasonal effect is becoming less severe for age group i.

- Does it look like all age groups have the same trend, with γi = γ₀ for all i?

- Do all age groups have the same cycle, with αpi = αp0?

- Does γ vary by region (negative in cold-climate regions?)

- Do the cycles, given by the α, vary by region or is a nation-wide cycle sufficient?

The above model is easier to interpret than the wavelet analysis. The quantity 1 − exp(120γ) is the change per decade in the seasonality effect, which could be reported with a 95% confidence interval.

Our current methods successfully answer the questions that the reviewer has posed. We have clarified our methods, especially those for Figures 12 to 17 and their respective figure supplements where applicable (referred to as the ‘second analysis’ above) in the revised manuscript and summarise the approach and its findings below.

- Is γi negative? If so the seasonal effect is becoming less severe for age group i.

- Does it look like all age groups have the same trend, with γi = γ₀ for all i?

We calculated change in percent difference between 1980 and 2016 by fitting a linear regression to the time series of seasonal differences in mortality within a year. We have reported the percentage point change in seasonality, with p-values, from 1980 to 2016 by age, sex, and cause of death (Figure 13 and respective figure supplements).

- Do all age groups have the same cycle, with αpi = αp0?

We calculated the timing of maximum and minimum mortality using centre of gravity analysis and circular statistics. We have reported the maximum and minimum values by age, sex, and cause of death (Figure 12 and respective figure supplements).

- Does γ vary by region (negative in cold-climate regions)?

We report percent difference in death rates between maximum and minimum months by region (Figure 18 and subsection “Statistical methods"). We have also calculated change in percent difference between 1980 and 2016 by region, shown in Author response image 1 for all-cause mortality, though not included in the main paper which is already dense with inclusion of a large number of results at the national level.

Author response image 1. Percent difference in death rates between the maximum and minimum mortality months for all-cause mortality in 2016 versus 1980 by sex, age group and region.

Author response image 1.

Open in a new tab

- Do the cycles, given by the α, vary by region or is a nation-wide cycle sufficient?

We calculated the timing of maximum and minimum mortality using centre of gravity analysis by region as well as nationally. We have reported the maximum and minimum values by age and sex for all-cause death (Figures 14-17 and subsection “Statistical methods").

Therefore, we have successfully answered all the questions posed by the reviewers with our current methods. The wavelet analysis allows greater flexibility when compared to the above model suggested because it does not require assuming only 6- and 12-month periodicity.

We have nonetheless implemented the above approach. We note that some the parameters of this model (e.g., the ρ, γ, and α terms) cannot be interpreted on their own, because the model includes involve interaction terms, for example (γ x α1i). In addition, the interpretation of γ is also dependent on the sign of α, as, for example, a negative value of α with a negative value of γ indicates the percentage difference between maximum and minimum mortality is increasing, whereas a positive value of α with a negative value of γ indicates that the percentage difference between maximum and minimum mortality is decreasing. For these reasons, interpretable results require predicting the outcomes in specific years and not simply the model parameters. We have done so for one of the outcomes presented in the paper, namely percent difference in death rates between maximum and minimum mortality months, with results presented in the table below. Conclusions were similar to those from the existing method.

We believe that our current approach has the dual advantage over the proposed method of using a more general and flexible framework and being more easily understandable. We welcome guidance on whether adding this additional method would improve the paper.

Sex Age (years) Percent difference 1980 Percent difference 2016
Male 0-4 11.9 6.9
Female 0-4 10.6 6.9
Male 5-14 56.2 19.4
Female 5-14 26.3 20.9
Male 15-24 39.1 10.8
Female 15-24 14.9 2.3
Male 25-34 17.3 7
Female 25-34 6.9 4.3
Male 35-44 4.9 4
Female 35-44 6.5 7.1
Male 45-54 10.9 7.9
Female 45-54 11.4 11.2
Male 55-64 13.7 11.7
Female 55-64 13.1 13.3
Male 65-74 17.6 14.5
Female 65-74 16.6 15.4
Male 75-84 24.3 20.2
Female 75-84 24.6 21.2
Male 85+ 35.2 25.5
Female 85+ 34.2 27.6
Open in a new tab

Author response table 1. Percent difference in death rates between the maximum and minimum mortality months for all-cause mortality in 2016 versus 1980 by sex and age group using the alternative method.

B) The approach is different to a number of those referenced (particularly the standard measure of excess winter deaths). It would be helpful to have some comment on this, particularly the generalisability to other studies. The focus has been on peak months in this analysis, but the standard assessment of excess winter deaths is to compare December to March months to the rest of the year.

We have avoided the term Excess Winter Deaths (EWDs) because the latter assumes higher deaths in winter than in the summer, and requires a priori decisions about which months to include in “winter” and “summer”. Our empirical results show that peak mortality not only varies from age group to age group or cause by cause, but also can in the extreme take place in summer, e.g., for young adult males. We have added a brief overview of EWDs, and why an empirical approach like ours is preferred to its strong assumptions (P. 19).

C) Appropriate rationale was given to the regional splits used in the analysis. However no description was given of the characteristics of these regional areas other than subsequent information on temperatures. As pointed out there could be a number of factors that are important. I would expect some reference to their differing characteristics.

A fair comment and addressed in the revised manuscript by providing an overview of the characteristics of the regions (subsection “Data”, second paragraph and Table 2).

4) Interpretation:

A) No table is presented with descriptive statistics, i.e. on the sample size of cases that fall into each disease category.

We have added a table of descriptive statistics (Table 1).

B) I may be misreading the paper (in which case it would be helpful to clarify so others do not make the same mistake), however the analysis in Figure 6 compares the difference in temperature experienced in those regions between the warmest and coldest months versus the% seasonal difference in death. In the Discussion this in contrasted with findings from Europe where countries with a more temperate winter have, paradoxically, higher rates of excess winter mortality. However this is not an appropriate direct comparison, regional extremes in variation were not looked at when comparisons between countries are made. This does not invalidate the comparison but it would be helpful to explicitly summarise if there are any differences between regions and examine what those are.

A fair comment and addressed in the revised manuscript in the Discussion. Specifically, the European papers that we had cited (Fowler et al., 2015; Healy, 2003; McKee, 1989) had based their comparison on annual mean temperature whereas ours was based on the temperature range. We have clarified this distinction in the revised Discussion, noting that the two temperature metrics are correlated (see Author response image 2).

Author response image 2. The relationship between annual mean temperature (used in European papers) and temperature range between maximum and minimum mortality months (used in our paper).

Author response image 2.

Open in a new tab

C) One option for this paper would be to be admittedly exploratory, avoiding the use of the word 'significant' and simplifying the analysis. Simple monthly averages and testing for the months having the same mean could replace the wavelet analysis. A second option would be to focus on a specific research hypothesis, explain carefully how the model estimates relate to this research hypothesis, and adjust the p-values for multiple testing.

As above, wavelet analysis, the subsequent steps of centre-of-gravity analysis to identify peak and minimum mortality months, and estimation of changes in seasonal mortality range are all formal analyses.

We now quote p-values in results, as opposed to setting an explicit threshold for significance (as derived by comparing against a threshold on the p-value), hence removing the need for correction for multiple testing.

Associated Data

    This section collects any data citations, data availability statements, or supplementary materials included in this article.

    Data Citations

    1. National Center for Health Statistics. 2017. Vital statistics (1980-2016) National Center for Health Statistics. dvs_data_release
    2. European Centre for Medium-Range Weather Forecasts. 2016. ERA-Interim temperature data (1979-2016) European Centre for Medium-Range Weather Forecasts. reanalysis-datasets/era-interim
    3. CDC Wonder. 2016. US Census Populations With Bridged Race Categories (1990-2016) CDC WONDER. bridged-race-population
    4. United States Census Bureau. 2016. County Intercensal Tables (1980-1989) United States Census Bureau. 1980s-county

    Supplementary Materials

    Transparent reporting form
    elife-35500-transrepform.docx (246.2KB, docx)
    DOI: 10.7554/eLife.35500.050

    Data Availability Statement

    The ERA-Interim temperature data are available at https://www.ecmwf.int/en/forecasts/datasets/reanalysis-datasets/era-interim. The US Census Populations With Bridged Race Categories data for 1990-2016 are available at https://wonder.cdc.gov/bridged-race-population.html. Pre 1990, the County Intercensal Tables are available at https://www.census.gov/data/tables/time-series/demo/popest/1980s-county.html. The Vital statistics data are available at https://www.cdc.gov/nchs/nvss/dvs_data_release.htm through a request to NAPHSIS (https://www.naphsis.org/).

    The following previously published datasets were used:

    National Center for Health Statistics. 2017. Vital statistics (1980-2016) National Center for Health Statistics. dvs_data_release

    European Centre for Medium-Range Weather Forecasts. 2016. ERA-Interim temperature data (1979-2016) European Centre for Medium-Range Weather Forecasts. reanalysis-datasets/era-interim

    CDC Wonder. 2016. US Census Populations With Bridged Race Categories (1990-2016) CDC WONDER. bridged-race-population

    United States Census Bureau. 2016. County Intercensal Tables (1980-1989) United States Census Bureau. 1980s-county

    Articles from eLife are provided here courtesy of eLife Sciences Publications, Ltd

    RESOURCES