METHODS
Data
We used weekly counts of all-cause mortality from January 2013 to
December 2020 tabulated by the National Center for Health Statistics
from the National Vital Statistics System. Data were stratified by age
group: ≤18 years of age, 18–49 years of age, 50–64 years of age,
65–74 years of age, 75–84, and ≥85 years of age; and on jurisdiction
of residence: the 50 states, District of Columbia, and New York City.
The data were tabulated on May 15, 2021. We used Morbidity and Mortality
Weekly Report weeks (MMWR weeks) to tabulate all-cause mortality. A MMWR
week begins on Sunday and ends on Saturday. The first MMWR week of a
year is the week with at least 4 days in that year.
Adjusting for secular trends
Because of population growth and aging, it is important to account for
long-term trends in the numbers of deaths occurring in the United
States. Because our methods relied on comparing recent data to
historical data, we needed to adjust for any secular trends in all-cause
mortality. First, we truncated the original time series by removing
weeks after September 28, 2019 to exclude both the 2019–2020 pneumonia
and influenza season as well as the COVID-19 pandemic. We decomposed
these historical data into three components using local regression
(LOESS): (1) a seasonal pattern, (2) a secular trend, and (3) the
residuals.13 When added together, the three components
of the decomposition are exactly the original data. Taken one at a time,
the seasonal pattern is exactly the same each year; the secular trend
slowly moves the seasonal pattern up and down; and the residuals are the
unexplained variation in weekly mortality. Because the secular trend
from this decomposition was defined only on the historical data, we
needed to extrapolate the secular trend to the end of 2020. So, we fit a
line to the secular trend from the decomposition using ordinary least
squares and extrapolated this line to the end of the original time
series. Next, we shifted this fitted line vertically such that its
minimum was zero on the domain of the original time series to prevent
taking the logarithm of a negative value later in our analysis. Finally,
we added this line to the original time series to adjust for the secular
trend over the entire study period.
Severity assessment
We used the adjusted time series for the severity assessment. We adapted
methods from Vega and colleagues to calculate intensity thresholds, a
part of their MEM.14 Because mortality in the United
States usually is at its highest during the fall and winter respiratory
virus season, we considered the historical data by season (October to
September), rather than calendar year. We used the 2013–2014 through
the 2018–2019 seasons and identified the 3 largest values of weekly
all-cause deaths from each season. Then, we assumed these values
followed a log-normal distribution. We calculated three intensity
thresholds corresponding to the 50th,
95th, and 99.5th percentiles of this
distribution: the median, 1.6 standard deviations above the median, and
2.6 standard deviations above the median. We denoted the intensity
thresholds as IT500, IT950, and IT995. These three thresholds defined
the four severity categories: “low severity”, “moderate severity”,
“high severity”, and “very high severity”. So, the peak weekly
mortality has a 50% a priori probability of being low severity,
45% of being moderate severity, 4.5% of being high severity, and 0.5%
of being very high severity. We used the intensity thresholds to assign
severity of each week from week 10 through week 53 (March 1, 2020
through January 2, 2021) to one of the four severity categories. We
repeated this analysis for the data stratified on age group and
stratified on state of residence.
Computer Software
We used “R: a language and environment for statistical computing” for
all computations.15 We also used the R package
“MMWRweek: convert dates to MMWR day, week and year” to manage
data.16