Can Photoperiod Predict Mortality in the 1918-1920 Influenza Pandemic?

Abstract

Amplitude of the seasonal change in day length increases with distance from the equator, and changes in day length markedly alter immune function in diverse nonhuman animal models of infection. Historical records of mortality data, ambient temperature, population density, geography, and economic indicators from 42 countries during 1918-1920 were analyzed to determine relative contributions toward human mortality during the “Spanish” influenza pandemic of 1918-1920. The data identify a strong negative relation between distance from the equator and mortality during the 1918-1920 influenza pandemic, which, in a multiple regression model, manifested independent of major economic, demographic, and temperature variables. Enhanced survival was evident in populations that experienced a winter nadir day length ≤10 h light/day, relative to those that experienced lower amplitude changes in photoperiod. Numerous reports indicate that exposure to short day lengths, typical of those occurring outside the tropics during winter, yields robust and enduring reductions in the magnitude of cytokine, febrile, and behavioral responses to infection. The present results are preliminary but prompt the conjecture that, if similar mechanisms are operant in humans, then they would be predicted to mitigate symptoms of infection in proportion to an individual's distance from the equator. Although limitations and uncertainties accompany regression-based analyses of historical epidemiological data, latitude, per se, may be an underrecognized factor in mortality during the 1918-1920 influenza pandemic. The author proposes that some proportion of the global variance in morbidity and mortality from infectious diseases may be explained by effects of day length on the innate immune response to infection.

Most human infectious diseases exhibit distinct seasonal patterns, peaking in the fall and winter in both hemispheres (Sakamoto-Momiyama, 1977; Dowell, 2001; Nelson, 2004). The causes of these seasonal cycles are not fully understood (reviewed in Lofgren et al., 2007). Changes in environmental day length exert potent immunomodulatory effects in nonhuman animals (Nelson, 2004). For example, in rodent models of innate immune stimulation, exposure to short (<10 h light/day) winter days decreases the magnitude of innate immune responses to bacterial (Bilbo et al., 2002) and viral infections (Baillie and Prendergast, 2008): proinflammatory cytokine responses are reduced, behavioral symptoms of infection (fever, anorexia, and cachexia) are decreased, and mortality from sepsis is markedly reduced (Bilbo et al., 2002; Prendergast et al., 2003, 2007). These non-specific behavioral and physiological symptoms of infection—such as anorexia and somatic wasting—contribute to morbidity and mortality. Mechanistically, the seasonal decreases in innate immune and behavioral responses to simulated infection are the result of photoperiod-driven changes in nocturnal melatonin secretion (Wen et al., 2007). In nature, immunomodulatory short winter day lengths occur only outside the tropics, and their magnitude and duration are augmented in proportion to distance from the equator. The bearing of photoperiodic changes in immune function on human immunity and epidemiology has not been clear, as urban exposure to bright artificial light functionally eliminates seasonal variation in photoperiod and melatonin secretion (Lewy et al., 1980; Wehr et al., 1995).

The global influenza pandemic of 1918-1920 provided a historically unique convergence of events and an opportunity to gain insights into the role of environmental day length in the human response to infectious disease (Patterson and Pyle, 1991). Artificial nocturnal illumination was far less prevalent during the 1918 pandemic relative to any subsequent pandemic interval (i.e., 1957-1958, 1968-1970, or 2009-2010). Accessible, inexpensive, bright artificial light became available only after the Industrial Revolution (Wehr, 2001); for example, in the decade before the 1918 pandemic, electric service was available in approximately 8% of US households (de Long, 1998). Because limited commercial airline service began in the 1920s (Davies, 1964), human migration was far more limited in 1918 relative to subsequent pandemic eras; thus, individual photoperiod exposure may be reasonably inferred from latitude of death. Lastly, modern antiviral treatments did not exist in 1918, nor did antibiotics effective in preventing secondary infections. Absent such treatment and prophylaxis, morbidity and mortality rates in 1918-1920 more closely reflected native human immune responses than they did during influenza pandemics that followed. The convergence of these epidemiological and technological events motivated an analysis of the most recently published mortality statistics from the 1918-1920 influenza pandemic (Johnson and Mueller, 2002) in an effort to gain insight into the relevance of changes in day length toward human health and disease during this pandemic.

MATERIALS AND METHODS

Mortality Data

Human mortality data were derived from Tables 1 to 5 of Johnson and Mueller (2002), which reflect the most recent peer-reviewed recalculation estimates of total mortality during the 1918-1920 influenza pandemic. Total mortality may include deaths due to harvesting and due to comorbid conditions hastened by influenza infection (e.g., pneumonia). Data are analyzed as log(death rate), where death rate equals recalculated deaths per 1000 people via analyses described in Johnson and Mueller (2002). Where ranges were given, the arithmetic mean of the range limits was used.

Table 1.

Results of Least Squares Regression Models of Single (Models 1-6) and Multiple Predictor Variables (Model 7) on log(Mortality) during the 1918-1920 Influenza Pandemic

Model	R 2	Independent(s)	n	β	SE	p Value	t Value
Model 1	0.425	Latitude	42	–0.652	0.004	0.000*	–5.438
Model 2	0.192	Per capita GDP	32	–0.438	<0.001	0.012*	–2.668
Model 3	0.264	Mean Ta	42	0.514	0.010	0.005*	3.790
Model 4	0.320	Ta range	42	–0.565	0.008	0.000*	–4.334
Model 5	0.030	Population density	42	–0.174	0.001	ns	–1.117
Model 6	0.032	Hemisphere	42	–0.178	0.098	ns	–1.146
Model 7	0.430	Latitude	32	–1.295	0.011	0.013*	–2.662
		Per capita GDP	32	–0.361	<0.001	ns	–1.702
		Mean Ta	32	–1.200	0.027	0.030*	–2.304
		Ta range	32	–0.228	0.012	ns	–0.901
		Population density	32	0.160	0.001	ns	0.825
		Hemisphere	32	0.222	0.098	ns	1.061

GDP = gross domestic product.

Latitude and Photoperiod Data

North and south latitude was determined based on the location of the major population center of each of the 42 countries for which recalculated death rates were available. Given that populations are distributed non-uniformly in space, no single value of latitude perfectly identifies the location of all individuals in a given country. The population centroid would be most accurate in this regard, but the available data from 1918 do not permit centroid calculation. Absent a calculation of population centroid, there exists the potential for the magnitude of a country's latitude range (i.e., the north-south distance of a given country) or the nonuniformity of the population distribution to constitute a source of error. The nadir photoperiod (interval, in hours, between local sunrise and sunset) at each major population center was determined for the years 1918-1920 using algorithms available at the US Naval Observatory (http://aa.usno.navy.mil/data/).

Economic Data

Per capita gross domestic product (GDP; purchasing power parity), defined as the value of all final goods and services produced within a nation in a given year, was adjusted for the relative cost of living and the inflation rates of each country, and data were obtained from The World Economy: Historical Statistics (Maddison, 2007). Economic data for 9 countries were unavailable during a 20-year interval prior to 1918 (Mauritius, Fiji, Guatemala, Mexico, West Samoa, Gambia, Nigeria, Cameroon, Kenya). GDP data from 1913 were available for 33 countries, whereas GDP data from 1917 (the year immediately prior to the start of the pandemic) were available for only 25 countries. Therefore, data from 1913 were used in all analyses.

Temperature Data

Temperature data for 1918-1920 were obtained via FTP (downloaded on April 30, 2009) from the Global Historical Climatology Network (GHCN-Monthly) database of the National Environmental Satellite, Data, and Information Service (www.ncdc.noaa.gov/ghcn/ghcn.html). This archive provides historical climatological data for thousands of land stations worldwide. Meteorological surface stations were selected based on their proximity to the population center in 1918 of each country included in the present report. For 3 countries (Guatemala, Cameroon, Gambia), data from 1918-1920 were unavailable. The nearest 2 to 3 continuous years available in the database (1931-1933, 1893-1894, and 1931-1933, respectively) were substituted for these 3 countries.

Geographic and Population Data

Area (in km²) of each country in 1918 was obtained from Table 1 of the United Nations Demographic Yearbook (Statistical Office of the United Nations, 1948). For hemispheric assignment, transequatorial countries were coded based on the location of their major population center. Population in 1918 was obtained from Tables 1 to 5 of Johnson and Mueller (2002). Population density was the quotient of population divided by area.

Historical Data Considerations

All estimates of pandemic mortality contain limitations (Cliff et al., 1986). Limitations include lack of registrations, misdiagnoses, and missing records. For example, the nature and prevalence of harvesting effects and comorbid infections may differ geographically. Indeed, successive accounts of influenza mortality from 1918-1920 are continuously upwardly revised (Johnson and Mueller, 2002). A recent report (Murray et al., 2006) questioned the quantitative rigors applied to previous estimates of pandemic mortality during 1918-1920, suggesting that the exclusive use of primary vital registration is required to develop high-quality statistical models for estimating future pandemic mortality. When such data are available, the report argues in favor of the use of the excess death rate method for the purposes of predicting future pandemic mortality above and beyond normal annual deaths (Murray et al., 2006).

However, the global coverage of high-quality vital registration data is geographically limited. For example, the Berkeley Online Human Mortality Database (Max Planck Institute for Demographic Research; http://www.mortality.org) contains detailed historical (1918-1920) vital statistics on 38 countries, none of which is located within the tropics. If a systematic effect of geography (whether mediated by latitude, temperature variation, humidity, poverty, etc.) impacts pandemic mortality, then geographically constraining the analyses may inadvertently limit reliable extrapolation beyond the regions represented.

The present analysis expanded geographic coverage by using data from Africa, the Americas, Europe, Asia, and Oceania (Johnson and Mueller, 2002). These recalculated mortality data were derived from a variety of sources, using excess mortality data when available but also using original records and extrapolations from subsets of the population within a given country (for detailed description, see Johnson and Mueller, 2002). This approach permitted inclusion of 15 countries from within the tropics, in a total data set of 42 countries.

It is beyond the scope of the present work to argue superiority for one or another source of mortality data on a pandemic that occurred far from recent memory. Ultimately, confidence in interpretations should be proportionate to confidence in the relative accuracy of the data collection. The use of the most current recalculated death rates available, encompassing the greatest geographic region available, may afford insights into the role of geophysical variables that cannot be examined in more limited data sets.

Nevertheless, in an effort to determine whether the death rate estimates used in Johnson and Mueller (2002) and the present report corresponded quantitatively with excess deaths mortality estimates used in another recent report (Murray et al., 2006), both sets of data were subjected to the same multivariate regression analysis. However, in this reanalysis, the data were restricted to countries present only in both data sets (i.e., nonequatorial countries only). The results were remarkably comparable (see Table 2), suggesting that the GDP and death rate data sets drawn from in the present report yield relations comparable to those using excess deaths mortality estimates, if equatorial countries are excluded.

Table 2.

Relative Contributions of Latitude and GDP to Pandemic Mortality in 1918-1920: Concordance of Murray et al. (2006) and Present Data

	Murray et al.	Present Data
Model 1: Economic predictor only
R2	0.473	0.609
n	27	23
β ± SE	–0.885 ± 0.187	–0.990 ± 0.167
p value	<0.001	<0.001
t value	–4.74	–5.94
Model 2: Absolute value of latitude and economic predictors
R2	0.482	0.590
n	27	23
Economic predictor
β ± SE	–0.967 ± 0.229	–0.979 ± 0.220
p value	<0.001	<0.001
t value	–4.22	–4.46
Latitude predictor
β ± SE	0.005 ± 0.008	0.0003 ± 0.004
p value	0.531	0.936
t value	0.64	–0.082

Comparisons between single and multiple regression model analyses of relations between latitude, economic variables,^a and 1918-1920 mortality^b based on all data in the present report (n = 23 countries) that overlap with published data from Murray et al. (2006). No countries between 23.4°N and 23.4°S latitude are included in these analyses.

^aMurray et al. (2006) data: per-head income derived from (Mitchell, 2003). Present data: gross domestic product (GDP)–purchasing power parity (PPP) derived from Maddison (2007).

^bMurray et al. (2006) analysis: pandemic excess mortality derived from the University of California at Berkeley, Max Planck Institute for Demographic Research, The Human Mortality Database (http://www.mortality.org) and via calculations described in Murray et al. (2006). Present analysis: recalculated pandemic death rate derived from Johnson and Mueller (2002).

Statistical Analyses

Linear regressions between independent variables and mortality were conducted using Statview software for the PC (SAS Institute, Cary, NC). Categorical analyses of nadir photoperiod and mortality used an analysis of variance (ANOVA) followed by Fisher's protected least significant difference (PLSD) tests. All data are depicted as means and standard errors of the mean. Differences were considered significant if p < 0.05.

RESULTS AND DISCUSSION

Death rates varied widely among countries, from 1.2 to 445 deaths per 10³ individuals, but a strong negative correlation was evident between distance from the equator and mortality (R² = 0.43; β = -0.652; p < 0.0001; n = 42; Fig. 1A; Table 1, model 1); this was observed among countries in both the Northern (β = -0.672; R² = 0.45; p < 0.0001; Fig. 1B) and Southern (β = -0.779; R² = 0.61; p < 0.01; Fig. 1C) hemispheres. Thus, a 20° shift in latitude toward the equator (e.g., from Montreal to Miami) would be predicted to be associated with a ~15% increase in mortality. The impact of latitude on mortality was comparable in the Northern and Southern hemispheres (hemisphere × latitude: β = 0.527; t = 1.84; p > 0.05), but in the tropics, where changes in photoperiod are attenuated, latitude did not predict mortality (β = -0.384; R² = 0.15; Fig. 1D).

Figure 1.

Worldwide mortality during the 1918-1920 influenza pandemic: relation to latitude. Panels show log-scaled mortality (deaths per 1000 people) among 42 countries. Linear regression of the death rate and the major population center latitude for (A) all countries for which data are available, countries segregated by (B) Northern and (C) Southern hemisphere, and (D) countries whose major population centers in 1920 were within the tropics. Dashed lines indicate 95% confidence intervals of the mean. Note ordinate axis is on a log scale.

Latitude correlates with several factors that may affect case fatality, including ambient temperature (T_a) and socioeconomic variables (poverty, medical care) (Murray et al., 2006; Lowen et al., 2007). In an effort to explain the relative contributions of these latitudinal covariates, least squares multiple regression analyses were performed using these variables as predictors and pandemic mortality as the dependent variable (Table 1). The regression model also examined the roles of population density and hemisphere. The model indicates that, in addition to latitude, economic (per capita GDP [Maddison, 2007]) and nonphotic environmental factors (mean annual T_a, T_a range) each individually explained global patterns of mortality (cf. Murray et al., 2006; Table 1, models 2-4); however, nearly half of the worldwide variance in mortality during the 1918-1920 pandemic was explained by latitude alone. In contrast, GDP and T_a alone each explained 19% and 26% to 32% of the variance in mortality, respectively. Moreover, in the multivariate model, which simultaneously considered multiple predictor covariables, latitude and mean T_a significantly explained mortality, whereas the coefficients for GDP and T_a range, which were significant in simple regression models, were not significant (Table 1, model 7). Thus, significant effects of latitude on 1918-1920 pandemic mortality were evident, independent of changes in all other variables investigated in this model.

Earlier work on this issue established a strong association between GDP and mortality during 1918-1920 but identified no association between latitude and mortality when considered in the context of GDP (Murray et al., 2006). The absence of detectable effects of latitude in prior work may arise from the omission of mortality data from countries located between the tropics. In contrast, data in the present report are derived from 42 countries, 15 (36%) of which were within the tropics (see “Historical Data Considerations”). If analysis of the present data is restricted to regions outside of the tropics, then it confirms conclusions reached by earlier reports (Murray et al., 2006; Table 2). Exclusion of data from tropical regions may omit a significant and informative source of variance and may have contributed to prior conclusions that pandemic mortality in 1918-1920 was unaffected by latitude.

When numerous environmental variables were controlled for (Table 1, model 7), effects of mean T_a on mortality were also evident. This relation may be mediated in part by T_a-driven constraints on hyper- and hypothermia and subsequent interactions between body temperature and survival of sepsis (Kluger et al., 1988; Romanovsky et al., 1996). In addition, T_a affects both transmission (attack rates) and symptom severity (case fatality) of human influenza viruses (Molinari et al., 2007; Lowen et al., 2007). Effects of T_a on influenza-associated mortality are undoubtedly complex, but the inclusion of T_a in multivariate models may permit more complete understanding of how local environments affect disease outcomes (e.g., Ballester et al., 1997). It is also potentially interesting to consider why the direction of the relationship between T_a and mortality changes between models 3 and 7 (Table 1). Clearly, when it is treated in isolation in model 3, the estimated effect of T_a (which includes the effects of any variables correlated with T_a) is positive. That is, lower T_a is associated with lower mortality, and higher T_a is associated with higher mortality. But once the effects of certain correlated variables are accounted for in model 7—par ticularly the pronounced effects of photoperiod—the partial correlation between T_a and mortality changes. Indeed, controlling for photoperiod, lower T_a actually seems to be detrimental to survival. This suggests that the zero-order correlation (model 3) may be spurious, allowing the beneficial effects of photoperiod to masquerade as effects of T_a.

In environments in which individuals had the opportunity to be exposed to seasonal changes in photoperiod, mortality during the influenza pandemic was attenuated in proportion to the absolute value of latitude. In light of the physical relation between latitude and the amplitude of seasonal changes in day length, these data raise the possibility that seasonal variations in day length, per se, may partially mediate the effects of latitude in the present analysis and may yield functionally significant changes in human immune function.

The phenomenology and mechanisms by which photoperiod alters immune function have been elaborated in several laboratory rodent models. Exposure to short day lengths alone is sufficient to attenuate symptoms of simulated bacterial and viral infections (Bilbo et al., 2002; Nelson, 2004; Prendergast et al., 2007). This symptom mitigation occurs as a result of decreased synthesis and secretion of IL-1β, IL-6, and TNF-α (Bilbo et al., 2002; Prendergast et al., 2003), combined with inhibited behavioral responsiveness to proinflammatory cytokines (Wen and Prendergast, 2007). In addition, short photoperiods trigger increases in circulating CD3+/CD4+/CD25+ Treg cells, which can inhibit proinflammatory cytokine production (Murphy et al., 2005; Prendergast et al., 2007). Moreover, during severe cases of innate immune inflammation, short days promote survival of lethal sepsis (Prendergast et al., 2003).

If exposure to short days indeed attenuates inflammatory responses in humans, then the question of when day length is acting to attenuate morbidity/mortality becomes of interest. In nature, the nadir photoperiod, the rate of change in photoperiod, and the annual duration of exposure to short days all covary, and thus a dissection of their relative contributions may not be possible using epidemiological data. Moreover, potential immunomodulatory effects of longer summer day lengths must be considered as well. In the laboratory, however, effects of short days on immune function are persistent, even in the face of intervening longer day lengths. For example, effects of short days on T cell–mediated immune responses endure for at least 5 months, even after subsequent exposure to long days (Prendergast et al., 2004). Thus, one possibility is that, despite the year-round prevalence of pathogens, the geophysically dictated unavailability of shorter day lengths in the tropics at any time of year may categorically deprive its inhabitants of the mere capacity for photoperiodic immunomodulation.

There is a strong positive correlation between cytokine production and the clinical severity of influenza viruses (Hayden et al., 1998; Fritz et al., 1999; Kaiser et al., 2001). Dysregulated and sustained increases in pro-inflammatory cytokine production (“cytokine storms”) result in hemophagocytosis, acute respiratory distress, and multiple organ dysfunction syndromes (Headley et al., 1997; Fisman, 2000; Huang et al., 2005) and are commonly considered the principal causes of the severe clinical presentations during pandemics (Kobasa et al., 2004; Guan et al., 2004). In some rodent models of infection, short days inhibit inflammatory responses to gram-negative, gram-positive, and viral mimetics, suggesting an omnibus effect of day length on diverse innate inflammatory responses (Bilbo et al., 2002; Baillie and Prendergast, 2008). Photoperiodic mitigation of proinflammatory cytokine signaling would presumably attenuate the severity of cytokine storms in humans. A significant limitation of the present analysis is that pandemic mortality rates may also reflect increased pathogen transmission rates, increased symptom severity, or a combination thereof, and the available data are unlikely to afford partitioning variance in mortality during 1918-1920 among these possible sources.

When the 1918-1920 influenza mortality data are categorized into hourly bins based on nadir winter day length, a clear enhancement of survival is evident in populations where, within a year prior to death, individuals may be assumed to have been inhabiting areas where photoperiods reached 10L or shorter (corresponding to ≥30° circle of N or S latitude; Fig. 2). Although the critical photoperiod for the inhibition of inflammatory responses in rodent models of innate inflammation is not known, a recent report indicated linear effects of photoperiods ranging from 15L to 9L on several aspects of immune and adrenocortical function in hamsters, with critical day lengths of 11L and 9L for effects of short days on T cell–dependent immune function and anti-inflammatory cortisol production, respectively, suggesting that progressively shorter day lengths may yield stepwise changes in some measures of immune function (Prendergast and Pyter, 2009). It is not known if the absolute value of the photoperiod (e.g., hours of light/day) interacts with the duration of photoperiod exposure (e.g., weeks of exposure) to affect immune responses.

Figure 2.

Nadir winter day length predicts mortality during the 1918-1920 influenza pandemic. Countries (n = 42) were categorized based on nadir winter day length rounded to the nearest hour. Sample sizes along abscissa indicate number of countries in each bin. *p < 0.002 vs. 11L and 12L groups.

Additional limitations of the present analysis include the necessity of reducing economic, temperature, demographic, and disease outcome data to single data points for the multiple regression analyses. Incorporating the continuous properties of these variables and dynamic changes therein would undoubtedly refine future analyses, as would including the timing of human mortality with respect to environmental conditions (photoperiod, temperature, humidity, altitude), energetic variables (nutritional status), and social variables (supportive and palliative care, medical education and infrastructure). One may endeavor to control for all reasonable independent variables in such analyses and, in the process, have the opportunity to evaluate interesting and plausible alternative factors, but ultimately, the analysis is limited by the data available. The present regression analyses do not indicate causality; however, the parallels that appear between correlational models and experimental data on changes in day length and symptoms of infection are noteworthy and may motivate or inform additional empirical efforts.

The effects of the environment on influenza transmission, symptom burden, and survival are undoubtedly complex and are unlikely to be accounted for by day length alone. But taken together, the present analysis identifies a strong correlation between latitude/photoperiod and mortality during the 1918-1920 influenza pandemic. This outcome raises the hypothesis that seasonal changes in day length may be of functional significance toward the outcome of some human infectious diseases. No study has directly addressed the role of seasonal photoperiod change or nightly melatonin secretion in the severity of infections in human populations, but the present analysis of epidemiological data, together with mechanistic clues from nonhuman animal models, suggest a potential biological relevance.