On Detecting Endemicity: Insights from All-Cause Mortality Patterns During Epidemic Transitions

During epidemic outbreaks, transitions to endemicity are often discussed, but rarely, if ever, defined. Analyzing all-cause mortality patterns may be useful.

As Vandenbroucke and Pearce write in this issue of AJPH (p. 593), all-cause excess mortality—rates exceeding expected values, agnostic to cause—provides perhaps the best and least biased measure of the most severe effects of an acute outbreak.

And yet, in the fullness of time, there is no such thing as an excess death, as all people die precisely once. These observations are not contradictions. Rather, they present an opportunity to understand features of a recent epidemic, and to aid in the detection of either the endemic phase of an outbreak, or the eradication of the novel threat. Here, I will discuss how by considering specific patterns of all-cause mortality, it can be determined when a pathogen has likely become endemic. In addition, these patterns can reveal important information about the cohort of excess decedents.

FACTORS IN ALL-CAUSE MORTALITY SUPPORTING ENDEMICITY

Several mortality-related measures can help determine if an endemic phase has been reached. While no single metric is definitive, the presence of one or more points toward endemicity:

• 1.The recoupling of observed and expected deaths. In an amply sized population, the correlations (weekly or monthly) between observed and expected deaths are reliably strong (Figure 1a, years 0–2). Acute outbreaks marked by sudden increases in mortality uncouple these correlations (Figure 1a, years 2–4). Endemicity is supported when recoupling of observed and expected deaths occurs (Figure 1a, years 4–7).
• 2.The resumption of typical mortality shares across demographics. Generally, the shares of all-cause mortality by age, sex, race/ethnicity, etc., are disparate, but stable. Outbreaks disrupt these norms. The return to preoutbreak mortality share patterns (including regrettable disparities) supports endemicity.
• 3.The resumption of typical variability across jurisdictions. Typically, variance in mortality by jurisdiction is stable. Outbreaks widen mortality dispersion across jurisdictions, as pathogens move through various regions. A return to preoutbreak variance, therefore, supports endemicity.
• 4.If novel pathogen-attributed deaths rise during times of increased disease incidence, but all-cause mortality does not, endemicity is supported (assuming decreases in other causes of deaths have not offset pathogen-driven increases).

FIGURE 1—

Mortality Dynamics Before, During, and After a Pandemic by (a) Observed and Expected Mortality Scenarios and (b) Different Mortality Scenarios

Note. Part a displays a hypothetical time series of mortality data over a 7-year period. The solid black line represents the observed mortality rate before a pandemic (years 0–2) and during a pandemic (years 2–4), The yellow dashed line shows the expected mortality rate in the absence of the pandemic (i.e., counterfactual mortality) starting in year 2. Starting in year 4, three postoutbreak scenarios are shown. Scenario A represents a chronic increase above expected mortality while scenario C represents a chronic decrease (“deficit mortality”) during the postacute phase of the outbreak. Scenario B represents a continuation of prepandemic expected mortality. Part b depicts a Kaplan-Meier survival curve under different scenarios. The yellow solid line symbolizes the original Kaplan-Meier survival curve (no outbreak). The green dashed line shows a scenario (similar to scenario C in part a) wherein a disproportionately greater increase in mortality occurred among individuals with lower life expectancies. The blue dash-dotted line depicts a scenario (similar to scenario A in part a) wherein a disproportionately greater increase in mortality occurred among those with higher life expectancies.

CESSATION OF EXCESS MORTALITY MAY NOT SUPPORT ENDEMICITY

Initially, I believed that the cessation of all-cause excess mortality or the appearance of deficit mortality might indicate endemicity. However, I wish to highlight that this may not always be so. Following an outbreak, three conditions may occur: all-cause excess mortality cessation, the appearance of deficit mortality, or chronic (though reduced) excess mortality (Figure 1a; years 4–7, scenarios B, C, and A, respectively). None of these conditions inherently supports endemicity, but instead, may carry information about the pandemic decedents (Figure 1b, Figure 2). Indeed, each of the three scenarios can occur absent the recoupling of observed and expected deaths, the resumption of usual expected shares, or typical variance ranges, as discussed. To understand why, let us consider the three postoutbreak conditions:

• 1.Excess mortality cessation (Figure 1a, scenario B). While appearing reassuring, if the epidemic disproportionately affected persons with shorter life expectancy, equilibrium has not been reached. Rather, revised rates of expected mortality should be lower than counterfactual expected rates. Therefore, the cessation of excess mortality (when observed deaths approximately matches expected deaths) might actually represent excess mortality, once pull-forward/healthy survivor effects have been considered.
• 2.The appearance of deficit mortality (Figure 1a, scenario C). After an outbreak, historically low death rates may appear. This, too, may be falsely reassuring. However, as above, if most excess decedents had short life expectancies, revised expected rates of postoutbreak mortality should be far lower. Therefore, what appears as deficit mortality might really be excess mortality.
• 3.The appearance of stable chronic excess mortality (Figure 1a, scenario A). The emergence of chronic excess mortality—i.e., low, relatively constant rates of excess mortality—is perhaps the most confusing possibility. (It also appears to be the condition we are seeing at this point in the COVID-19 experience.) While chronic excess mortality might seemingly argue against endemicity, continuous but relatively stable excess mortality can be consistent with endemicity (for reasons not dissimilar from those mentioned previously, though that may seem surprising). There are at least two possible explanations to consider. First, if the pathogen remains dangerous enough to the survivor cohort that it poses an ongoing (albeit lower) risk, chronic excess mortality may occur. That is, any pull-forward/healthy survivor effects are offset by the pathogen’s continued presence. Second, chronic excess mortality can appear (despite outward appearances) when the mortality distribution density during the acute outbreak has, on average, been disproportionately concentrated among the healthier half of the population (Figure 2a; the curve shows homogeneous excess mortality across the risk spectrum compared with Figure 2b, where changes are heterogeneous across the risk spectrum). This pattern would be observed if risks taken in the population correlated to perceived health except at the far left of the curve shown in Figure 2b; restated, healthier people (the right half of Figure 2b) experienced greater exposure to the pathogen (owing to higher risk tolerance) during the outbreak. However, persons with extremely high mortality risks (i.e., those with high risk of mortality during years 0–3) may have been unable to avoid pathogen exposure; for example, persons living in long-term care facilities, or those who require frequent medical visits. This scenario could explain why nursing home residents experienced high mortality rates during the early COVID-19 outbreak, but today’s survivor cohort may (seemingly paradoxically) actually be less healthy on average than the prepandemic population. If mortality rates on the far right of the graph were, say, 8-fold below the rates on the left (e.g., 1000 vs 8000 deaths per 100 000 population), but associated years of lost life were, on average, 10-fold greater (e.g., 15 years vs 1.5 years), the survivor population would have lower life expectancies. A scenario like this is rarely considered but seems likely.

FIGURE 2—

Pandemic Mortality Dynamics of a Cohort of 80-Year-Old Persons (a) Overall and (b) With Risk-Specific Pandemic-Associated All-Cause Mortality Odds Ratios Shown as Density Under the Curve

Note. Part a shows a normal distribution mortality curve. The mean life expectancy for the cohort is represented by the central white dashed vertical line. The density (ρ) of shading represents the relative contribution to increased mortality during an outbreak, across the life expectancy spectrum. The homogenous shading indicates that, in this instance, any changes in mortality are evenly distributed across the risk pool. Part b shows a similar curve as the left, but with altered relative contributions to increased mortality. In part b, the average density to the right of the mean (vertical dashed line) is greater than the density (ρ) to the left of the mean (despite the local area with the greatest density appearing at the far left).

ENDEMICITY VS ERADICATION

Without further analysis, conditions supporting endemicity might be indistinguishable from conditions consistent with eradication. However, if eradication has occurred, the Figure 1 scenario C and A curves should eventually merge with the scenario B curve (sooner in scenario C and later in scenario A, reflecting the differences in years of lost life per excess death). An equilibrium would only be reached once the area under the curve from the acute phase (Figure 1a, years 2–4) equals either the area between the curves of scenario A and B or the area between the curves of scenario B and C during the postacute phase. An alternative way to visualize this would be to observe a smaller-than-expected population after an outbreak (Figure 1b) returning to normal levels either four years (pull-forward effect) or 11 years later.

Meanwhile, in scenario B, distinguishing between eradication and an endemic threat that has essentially been neutralized (by immunity or treatments that abolish mortality) is challenging. Combining these observations with viral surveillance (e.g., wastewater monitoring), however, would be definitive.

Alternatively, if a pathogen has become endemic (marked by low but persistently elevated mortality), the curves of scenarios A and C would never merge with that of scenario B. Here, the life expectancy of the survivor cohort would eventually be found to have increased (scenario C) or decreased (scenario A) compared with the preoutbreak era population.

CONTEXTUALIZING EXPECTED DEATHS (THE “NULL VALUE”)

The general usefulness of excess mortality turns on the reliability and interpretation of expected deaths (the “null value”). Let us explore this during the three outbreak phases:

• 1.Prior to outbreaks, long-term analyses of observed deaths define the bounds of predictability. Identifying how much variance (including seasonal variation) exists at baseline can help epidemic teams determine if the magnitude of a sudden mortality increase is cause for concern—that is, local thresholds that divide signal and noise. Is a weekly increase over expected deaths of 2% meaningful? Is 10%? Models can easily determine thresholds for mortality that are two, three, or many standard deviations above historical means. These thresholds depend on the population size, baseline mortality rates, and the general stability of preoutbreak conditions. The study of preoutbreak all-cause mortality, therefore, can establish thresholds useful for threat detection (adjusted for lag, if needed). While this work was not common before COVID-19, the rapid assembly of models identified that sudden increases in all-cause mortality that accompanied the initial outbreaks were massive outliers.
• 2.During outbreaks, knowing the number of expected deaths permits context (i.e., a counterfactual) for the number of recorded deaths. This allows an assessment of the magnitude of the problem. Nor is lag an insurmountable obstacle for using all-cause mortality this way. In Massachusetts, more than 98% of weekly deaths are routinely known to the Registry of Vital Statistics (and its partners) within three days of the previous week.
• 3.After outbreaks. The return to mortality rates similar to preoutbreak levels can, as mentioned previously, be helpful in determining characteristics of the endemic phase and the survivor cohort.

OTHER OBSERVATIONS REGARDING EXCESS MORTALITY

When considering excess mortality, unit choices are often overlooked. For instance, the number of excess deaths may seem small in populations with low baseline mortality rates, which may obscure important implications. For example, if mortality among a cohort of 30-year-old persons increases by 40% (from 100 to 140 among 100 000 people), only 40 excess deaths would be measured. However, assuming a life expectancy of 80 years, the years of lost life would amount to 2000. Meanwhile, a 10% increase among a cohort of 80-year-old persons (from 1000–1100 per 100 000) with prior life expectancies of eight years would result in 100 excess deaths, but only 800 years of lost life. Generally, the lower the baseline mortality rate, the greater the years of lost life. Deciding whether to report the number of excess deaths, rate (per capita), or observed-to-expected ratio dramatically influences the insights; these metrics have distinct meanings.

Additionally, while age standardization is useful in comparing outcomes across jurisdiction, the practice obscures ground conditions. Age-stratification may be more appropriate. For example, a region with many older persons might need to take greater precautions during outbreaks. Age standardization can give the false impression that such a region performed “as well” as one with a younger population, when, in fact, that was not so. Choices regarding measurement and reporting should be made carefully.

CONCLUSIONS: EMBRACING A CONTRADICTION

While excess deaths may not, in the final analysis, truly exist, early deaths do. Indeed, 99.9% of those who died of 1918 H1N1 would be dead by now anyway. Nevertheless, the tens of millions who perished in 1918 to 1919 were deprived of unrecoverable years of lost life.

Herein, I have proposed a paradigm for the detection of the endemic phase of an outbreak with respect to mortality. However, endemicity has different meanings. Declaring a pathogen endemic with respect to mortality alone does not imply that morbidity has been neutralized. Furthermore, endemicity may prove transient, especially with a rapidly mutating virus. Still, without a specific framework, there is the risk that subjective declarations of endemicity will be made once we are exhausted of a pathogen, but well before it has exhausted its most profound effects upon us.