Conceived and designed the experiments: RED CER KBE. Performed the experiments: CER. Analyzed the data: CER RED KBE. Wrote the paper: RED CER KBE.
The authors have declared that no competing interests exist.
The substantial winter influenza peak in temperate climates has lead to the hypothesis that cold and/or dry air is a causal factor in influenza variability. We examined the relationship between cold and/or dry air and daily influenza and pneumonia mortality in the cold season in the New York metropolitan area from 1975–2002. We conducted a retrospective study relating daily pneumonia and influenza mortality for New York City and surroundings from 1975–2002 to daily air temperature, dew point temperature (a measure of atmospheric humidity), and daily air mass type. We identified high mortality days and periods and employed temporal smoothers and lags to account for the latency period and the time between infection and death. Unpaired
It is well known that intra-annual mortality exhibits a pronounced winter peak in locations with seasonal climates
Recent research examining climatic influences on influenza transmission has galvanized interest in this topic. Airborne transmission of Influenza A/Panama virus between guinea pigs was more likely at low temperatures and relative humidities
These recent studies
A variety of theories exist as to how weather and climate might exert some influence on influenza seasonality. Low temperatures enhance viral stability
From a micro-physical perspective, there is evidence that both droplet size and transmission mode depend on ambient environmental factors
Theories proposing that factors other than weather/climate are responsible for influenza seasonality include cycles in viral interference
We examine the hypothesis that cold and/or dry weather enhances human pneumonia and influenza (P&I) mortality through a retrospective study of daily mortality in New York City and environs from 1975–2002. We hypothesize that periods with colder and/or less humid conditions exhibit excess P&I mortality for a period of time following those climatic conditions. Our research differs from recent work on this topic that examined the timing of the influenza season onset over large geographic areas
We selected New York City for this study for several reasons. Our study examines daily mortality, and statistical robustness is enhanced when the daily sample size is sufficiently large. Weather obviously has a high spatial variability, so it is important that the observed weather be representative of the environmental conditions likely experienced by the decedents. In addition, New York City's mid-latitude location provides a high degree of both interannual and intra-annual variability in weather and climate, so this variability provides a wider range of sample conditions. Thus, New York City has both a large enough population to provide a consistent daily mortality signal while the population density is high enough that the weather observed at a single station is sufficiently representative of conditions experienced throughout the metropolitan area.
We conducted a retrospective cohort study of pneumonia and influenza (P&I) mortality of residents of the New York City metropolitan area. Daily frequencies of P&I mortality were tallied from National Center for Health Statistics archives for the New York City Consolidated Metropolitan Statistical Area (which, as defined in the year 2000, includes 30 counties in New York, New Jersey, Connecticut, and Pennsylvania). This period of record spans three revisions of the International Classification of Diseases (ICD) codes (
REVISION | DATES APPLICABLE | CODE |
8th | 1975–1979 | 470–474, 480–484 |
9th | 1980–1998 | 480–484, 487 |
10th | 1999–present | J10–J16, J18 |
In the National Center for Health Statistics mortality files that we used for this research, all information that could allow an individual to be identified has been removed. This research utilized only mortality counts for a large metropolitan area. These de-identified counts are stored in governmental archives for the purposes of retrospective research; because all personal identifying information is redacted, consent is not required. Thus, this research is exempt from IRB review under the auspices of Title 45 Part 46 exemption category 4.
Pneumonia or influenza must be listed as the primary cause of death to be included in this analysis. These diseases are commonly combined as an endpoint because of specific challenges associated with influenza. First, the number of deaths attributable to influenza is difficult to estimate directly because of a lack of virologically-confirmed infections. Second, many influenza-associated deaths occur from secondary complications when influenza viruses are no longer detectable by laboratory means
Daily deaths were aggregated into ten age groups (0–4, 5–14, 15–24, 25–34, 35–44, 45–54, 55–64, 65–74, 75–84 and >84) and standardized via direct standardization
Examination of the P&I mortality time series exhibits obvious temporal discontinuities that are exactly coincident with the dates of ICD revision code changes. We removed this artifact by converting each day's P&I mortality to a z-score by dividing the mean departure by the standard deviation separately for each of the three relevant ICD periods (
a) (top) Daily age-standardized pneumonia and influenza mortality time series (deaths per million; June, July and August have been deleted). The relevant periods for the International Classification of Diseases (ICD) are identified by a thick vertical line; b) (bottom) Resulting mortality time series after removing the seasonality and converting to z-scores for each ICD period. Vertical dividers identify influenza seasons (September–May) with the year assigned to the January–May period (i.e., December, 1979 is in the 1980 flu “season,” labeled as “80” on the x-axis).
Mortality data were smoothed using a 17-day leading moving average (e.g., mortality on January 1 is the mean from January 1–17). This smoother was selected after testing a variety of filter lengths and based upon prior research
We examined both daily P&I mortality “events” and longer mortality “episodes.” “Events” are days with (smoothed) mortality at least one standard deviation above the long-term (smoothed) mean for that date. This z≥1 criterion was chosen because the frequency distribution of smoothed mortality is positively skewed with the tail beginning at approximately one standard deviation. After smoothing, there is an obvious tendency for high mortality events to cluster into prolonged periods when the z≥1 threshold is exceeded (
The time series is
SEASON | START | END | DURATION(DAYS) | TOTALDEATHS | AVERAGE DEATHS/DAY |
1976 | 19-Jan | 28-Mar | 70 | 144.58 | 2.07 |
1978 | 14-Dec | 2-Feb | 51 | 66.84 | 1.31 |
1980 | 16-Jan | 21-Mar | 65 | 82.67 | 1.27 |
1981 | 22-Nov | 6-Feb | 77 | 137.21 | 1.78 |
1985 | 15-Jan | 25-Feb | 42 | 44.54 | 1.06 |
1986 | 27-Feb | 26-Mar | 28 | 28.60 | 1.02 |
1988 | 12-Mar | 11-Apr | 31 | 28.55 | 0.92 |
1990 | 27-Dec | 19-Feb | 55 | 66.19 | 1.20 |
1992 | 31-Dec | 21-Jan | 22 | 22.61 | 1.03 |
1993 | 30-Jan | 23-Mar | 53 | 65.68 | 1.24 |
1999 | 5-Jan | 4-Mar | 59 | 78.26 | 1.33 |
2000 | 16-Dec | 1-Feb | 47 | 51.29 | 1.09 |
Deaths are age-standardized deaths per million in z-score units. Total deaths and average deaths per day include the entire time period between the start and end of the episode.
To summarize the outcome data treatment, after age standardization the daily mortality data were deseasoned to remove the large influence of season on respiratory infection and converted to z-scores to adjust for discontinuities related to ICD coding. These data were then smoothed using a 17-day leading smoother to account for the inherent lag between infection and mortality. Days or periods with z-scores>1 were identified as mortality “events” or “episodes,” respectively (
Hourly climate data from La Guardia Airport, New York, were retrieved from National Climatic Data Center archives. We utilize 1200 and 1900 Universal Time Coordinate (UTC) air temperature (T) and dew point temperature (Td) to approximate the typical times of the warmest and coldest hours of the day (7 or 8 a.m. and 2 or 3 p.m. local time). The dew point temperature is the temperature at which water vapor begins to condense via cooling at constant pressure. We use dew point as a measure of the amount of moisture in the air because, unlike relative humidity, it is independent of air temperature
In addition to dew point temperature, we employ an air mass classification, which has the advantage of incorporating a variety of weather variables into a single nominal variable. Specifically, we utilize the Spatial Synoptic Climatology (SSC)
The temperature and dew point time series were converted to z-scores to remove seasonality and then smoothed using a centered 5-day moving average filter. This filter length was employed after examining various options because it represents a balance between high frequency weather events and more long-term (monthly to seasonal) trends.
The seasonality of the SSC air mass types was removed by comparing the presence (coded 1) or absence (coded 0) of each air mass type on each day of the year to the long-term average frequency. For example, if Moist Moderate air was present on average 30% of the time on January 1 over the period of record, then its occurrence on January 1, 2000 would result in a value of +0.7 for that day. These daily anomalies were then converted into continuous variables using a centered 7-day moving average filter for each SSC category.
In summary, raw dew point observations were first de-seasoned by conversion to z-scores and then smoothed using a 5-day filter. Daily air mass frequencies were converted from a nominal to a continuous variable by first adjusting for the long-term frequency on each day and then smoothing those frequencies using a 7-day moving average (
A series of
N′ = adjusted degrees of freedom
P = lag one temporal autocorrelation.
N′ was then adjusted again based on the length of the smoother employed to determine the final effective sample size. For these and all other tests, a Type I error rate of 0.05 was employed and Levene's test for equality of variances was used to determine if pooling of the samples was required.
The following tests were performed:
smoothed temperature, dew point, and air mass frequency, lagged 17-days, between mortality events (z≥1) and non-events (z<1) using an unpaired two-sample t-test (
same as in 1 for unsmoothed temperature and dew point temperature (to determine if a strict 17-day lag exists); and
same as in 1 but using a one-sample t-test (to account for the possible influence of disparate sample sizes between groups).
For each high P&I mortality episode, we calculated the duration (in days), the summed total mortality over the entire episode, and the average daily episode mortality (
Given the relatively small number of episodes, we used bootstrapped regression analysis to generate a robust estimate of the regression coefficients. Based on the initial full sample, data sets of the same size were generated by randomly sampling variable pairs, with replacement, and estimating the regression parameters from that sample using ordinary least-squares. This procedure was repeated 10,000 times and the resulting suite of regression coefficients was examined to determine if the 2.5 percentile and 97.5 percentile observations were of the same sign. If so, the regression slope was deemed to be statistically significant
In the daily analysis, 1200 UTC dew point temperature was significantly lower for events than for non-events (p = 0.003), a result that is consistent with the hypothesis that drier conditions are related to enhanced P&I mortality. However, 1900 UTC dew point was higher during events (p = 0.036), a result that contradicts the underlying hypothesis.
When this test was repeated without smoothing the weather variables, only 1200 UTC dew point was significant (p = 0.028;
2-sample Smoothed | 2-sample Unsmoothed | 1-sample | |||||
Event | Non-Event | p | Event | Non-Event | p | p | |
1200 UTC T | −0.114 | −0.025 | 0.068 | −0.120 | −0.026 | 0.150 |
|
1200 UTC Td |
|
|
|
|
|
|
|
1900 UTC T | −0.012 | −0.028 | 0.725 | −0.012 | −0.029 | 0.767 | 0.783 |
1900 UTC Td |
|
|
|
0.056 | −0.032 | 0.139 | 0.181 |
Dry Moderate | 0.013 | 0.005 | 0.984 | n/a | n/a | n/a | 0.239 |
Dry Polar | −0.000 | −0.008 | 0.277 | n/a | n/a | n/a | 0.994 |
Dry Tropical | 0.001 | 0.002 | 0.998 | n/a | n/a | n/a | 0.561 |
Moist Moderate |
|
|
|
n/a | n/a | n/a |
|
Moist Polar |
|
|
|
n/a | n/a | n/a |
|
Moist Tropical | 0.000 | 0.001 | 0.691 | n/a | n/a | n/a | 0.931 |
Transition | −0.009 | 0.001 | 0.484 | n/a | n/a | n/a | 0.153 |
Air mass analysis could not be run without smoothing (n/a = not applicable). Results with p≤0.05 are shown in bold. Mean values for events and non-events are air temperature (T) and dew point temperature (Td) departures from the long-term daily mean in z-score units. Air mass values are mean frequencies based on a 7-day centered moving average filter. The z-score values for events in the 1-sample test are the same as in column 2.
Because the high number of non-event days
For the daily SSC analysis, lower frequencies of moist moderate (MM) air (p = 0.019) and higher frequencies of moist polar (MP) air (p<0.001) occurred 15–19 days before high mortality events (
Temperature, dew point temperature, and air mass frequencies were examined 17 days prior to and throughout each of the 12 high P&I mortality episodes identified from 1975–2002. There is a statistically significant negative relationship between episode duration and mean 1200 UTC dew point (r = −0.61, p<0.05;
a) (left) Total episode duration (days)
Dew point temperature is commonly used by atmospheric scientists to measure humidity because it is relatively invariant to pressure and temperature changes and thus is a conservative quantity. In New York City from 1975–2002, periods with high P&I mortality were preceded 2–3 weeks by periods with low morning dew points. Furthermore, for the 12 high mortality episodes identified in that period, morning dew point was negatively correlated with both episode duration (r = −0.61) and total episode mortality (r = −0.56). This finding of a linkage between dry air and influenza mortality is consistent with the results of recent research
The association between high frequencies of Moist Polar air masses prior to high mortality events is consistent with the dew point results. The average morning dew point in Moist Polar air in New York City is lower than for any air mass other than Dry Polar (
AIR MASS | FREQUENCY | RANK | MORNING Td | RANK |
|
25.9 | 2 | −9.3 | 1 |
|
7.3 | 4 | −2.0 | 2 |
|
31.2 | 1 | −1.5 | 3 |
|
13.6 | 3 | 3.5 | 5 |
|
1.7 | 6 | −1 | 4 |
|
6.9 | 5 | 9 | 6 |
We also identified a significant relationship between low Moist Moderate air mass frequencies 3–4 weeks before mortality episodes. In an effort to understand this association, we calculated the correlation between Moist Moderate frequencies and the other air mass types during that period. Moist Moderate is negatively correlated with the two driest air masses—Dry Polar (r = −0.47) and Moist Polar (r = −0.31) (
The lack of a direct Dry Polar relationship is surprising, as Dry Polar air masses exhibit the most extreme combination of cold air and low humidity. Dry Polar is far more common than Moist Polar in New York winters, however, so its high daily frequency during the influenza season limits the likelihood of identifying an underlying relationship. It might be more fruitful to examine an extreme cold, dry subset of Dry Polar air masses to identify the coldest and driest days.
In New York City, high P&I mortality periods within a given year were preceded by multiple day periods with unusually low temperature and humidity. Over the 28-year period of this study, we identified 12 episodes of high P&I mortality and found that both the total mortality occurring during each episode and duration of each episode were inversely correlated with the average morning dew point temperature prior to and during the episodes. These results support the burgeoning hypothesis that unusually cold dry air enhances the airborne transmission of influenza virus.
The exploratory nature of this analysis was necessitated by the lack of an underlying theory of influenza seasonality, socio-behavioral factors, and inherent variability in disease transmission and virulence. The time between infection and a resulting mortality event (i.e. “latency”) varies between individuals depending on age, overall health, co-morbid conditions, and other factors. Thus, lags must be estimated to best fit the overall data structure. Similarly, the high frequency variability in the variables requires some smoothing to elucidate relationships, and the selection of appropriate smoothers is somewhat subjective. Nevertheless, our findings are consistent with several others. For example, there is evidence supporting a two-week lag between rising influenza virus and pneumonia mortality
For this study, P&I mortality was used to characterize the influenza time series in New York City. The limitation of this method is the potential for confounding as P&I mortality includes mortality from infections other than influenza. In addition, in non-pandemic years, P&I mortality is skewed by the extremes of age. This limitation is unlikely to be a major contributor in this study as 90% of influenza-related deaths involve persons over the age of 65 during seasonal epidemics
We chose to focus on New York City because the large population (and thus large daily P&I mortality rate) enhances statistical robustness, and New York weather is highly variable owing to its midlatitude, coastal location. These results should be confirmed using a similar methodology in other cities worldwide to determine if the humidity-influenza linkage is pervasive. It would be particularly interesting to determine how these relationships evolve in subtropical or tropical climates where the P&I mortality seasonality is more muted or nonexistent.
It is likely that the underlying causes of influenza seasonality are multi-factorial, and we suspect that weather is but one of those factors. A predictive model for P&I mortality based on weather alone would likely be unsuccessful in accounting for most of the short-term influenza variability. Nevertheless, our results confirm recent emerging hypotheses of a relationship between cold, dry air and influenza transmission or virulence
(TIF)
(TIF)
(TIF)