|One unique feature of the pandemic of 1918 in the US was that mortality figures in different cities showed tremendous variations (up to 4-fold). This now-famous chart shows the difference between Philadelphia and St Louis.
2 recent studies published in PNAS analyzed the data from 23 cities in detail to determine if social distancing measures accounted for such differences, if so by how much, and what lessons we might draw for today.
The first one, Public health interventions and epidemic intensity during the 1918 influenza pandemic Hatchett, Mecher, and Lipsitch, compared the timing, number and type of interventions in 17 cities in relation to their peak and overall excess mortality rates. In all, 19 different types of interventions were studied during the fall wave of 1918.
First, a couple of (slightly) technical terms to explain: since we do not have actual data on how many people died as a result of the pandemic virus, one way to estimate the death rate in the population would be to use the excess deaths due to pneumonia and influenza (excess P&I deaths) compared to what the normal death rates would be during that time of the year. ie we assume excess PID = mortality rate due to pandemic
Then, as the pandemic progressed, we can also measure the total no of deaths up to a specific time, ie cumulative deaths. Thus, the CEPID (cumulative excess P&I deaths) becomes a measure of the total number of people who had died as a result of the epidemic up to that point.
If we assume that the CFR remained the same, then CEPID is also a reflection of the AR (attack rate or no of people infected) up to whatever point we are referring to.
Finally, if we assume the epidemic unfolds initially in similar ways in different cities, we can also compare CEPID between different cities, and assume that the same CEPID would represent the same phase of an unmitigated epidemic wave. ie use the same CEPID as a measure of timing eg CEPID at 20/100,000 population would reflect the same point of an epidemic wave in different cities, and would be earlier in time than 30/100,000 population in whatever city you are considering.
OK, that's all the technical stuff you need to know. Now onto the paper.
- In comparing the various parameters, the study showed that the strongest correlation was between the timing when interventions were introduced and lowered peak mortality. (Now, that was painless, wasn't it?) ie the earlier the set of interventions were implemented, the lower the peak death rate.
- Secondly, there was a correlation between the number of interventions and reduction in peak deaths.
- These two in combination give the following result: cities with 4 or more interventions before CEPID reached 20/100,000, compared to those with 3 or less, had death rates of 65 as opposed to 146/100,000.
- A weaker correlation was obtained for total deaths, at 405 vs 551/100,000.
- Some interventions were effective, some were not. Cities that closed school, theaters, and churches (often announced together) before CEPID reached 30/100,000 had peak deaths at 65-68 vs 127-146 /100,000 for those that didn't.
- Interventions that did not show significant correlations: closure of dance halls, other closures, isolation of cases, bans on public funerals, and making influenza notifiable.
- Some cities experienced second waves. This gave some very significant insights. Cities that had the most effective interventions ie most reduction in peak, in the first wave were most likely to have bigger second waves. They also tended to have them sooner than those that did not have very effective NPIs. Most crucially, no city experienced a second wave while NPIs were in force.
What do these results tell us?
Multiple NPIs applied early, with the emphasis on early, were correlated with reduced peak mortality and flattening of the epidemic curve.
The benefits only applied while the NPIs were in force. The epidemic came back when NPIs were lifted, and slowed down again if the NPIs were re-applied.
There was a smaller effect on reduction of total mortality.
Now, we shouldn't take researchers on faith. So we should ask this question, how do we know that the differences in death rates were due to NPIs and not due to other factors such as differences in virulence either between cities or over time?
There are several indications:
- they removed any variations in death rates by dividing the peak deaths by the median death rates over the epidemic of that particular city, so that cities that might have had an exceptionally high or low death rates would have that effect canceled out in this `normalized' peak death rate, which, after the adjustments, still showed the same degree of correlation.
- Not all interventions resulted in reductions in death rates.
- The correlation between the stage of the epidemic ie CEPID was stronger than with the actual calendar date of the onset, showing the effects were not due to attenuation of the virus over time, even though some attenuation did happen by January 1919.
- The presence of second waves only in those cities with effective NPIs shows that the reduction in the first peak were due to reduction in transmission and no of cases, and not due to a milder virus. Indeed, the most significant effect is that no second waves were observed in any city when NPIs were in force, which tells us that the second waves were caused by insufficient number of people having immunity in the city.
Onto the second paper, The effect of public health measures on the 1918 influenza pandemic in U.S. cities by Bootsma & Ferguson. This one is rather more esoteric, but when de-constructed, is still fairly easy to understand and similarly fascinating.
The first thing is, they analyzed the data and came to more-or-less the same conclusion as Hatchett et al, that multiple `imperfect transient interventions' when applied early was correlated with reductions in peak deaths, and the strongest correlation was with the timing of intervention, ie earlier interventions gave bigger reduction in peak deaths. There was also a weaker correlation with total deaths ie total epidemic size.
Having gotten an idea of what NPI's might do, the next thing they did was to take the historical data of how the epidemic unfolded in 1918 in the various cities, and use models of disease transmission and effects of interventions to try and predict what the outcome might have been, comparing the two to see how well the results fitted what actually happened.
The answer? There was very good fit! Showing that the correlations that we hypothesize as possible cause and effect were reproducible in a theoretical model, as the following example for St Louis shows.
The 1918 data is shown as purple crosses, the red curve is the predicted outcome from modeling, which in most cases fit the actual outcome pretty closely. The dark green curve is what it would have been if there were no interventions, the horizontal light green lines show the timing of intervention, with the height of the line showing their effectiveness. You can see when the interventions were started by looking at the beginning of the horizontal green line in relation to the epidemic curve. Notice how for Philadelphia the actual outcome was worse than the model predictions and very close to the worst case scenario, ie no mitigation at all.
In addition, the paper also explored some very important concepts. In order for an epidemic to be over, there needs to be enough immunity in the population, ie herd immunity. In theory, herd immunity is achieved when the proportion of people infected (ie AR) = 1-1/R0. The 1918 fall wave in the US was estimated to have an R0 of 2, which in theory would have meant that the epidemic would stop when 50% of people were infected. But, in reality, due to the time needed for those infected to acquire immunity, uncontrolled epidemics tend to overshoot, so that completely unmitigated epidemic would result in AR of 80%!
Now, transient NPI's, as we said, result only in transient reduction of transmission. When NPI's are very effective, the no of infections are so low that when the NPI's are lifted, there is still a large proportion of people without immunity, and the epidemic takes off again in an uncontrolled manner ie with overshoot, resulting paradoxically in an AR of say 65% (see chart above).
In principle, the most you can hope for, in the absence of vaccines, is to achieve the baseline AR of 50%, and the way to do that is to maintain NPI's just sufficiently to have some ongoing transmissions that would build immunity, and maintain NPI's for long enough until 50% immunity is achieved, which could take a fairly lengthy period of interventions.
Finally, using the models, the study estimated that in 1918, for those cities with the most successful NPIs eg St Louis or San Francisco, if they had been able to maintain their NPIs for however long it took to achieve herd immunity, they would have resulted in close to 40% reduction in total mortality, instead of the 10% (St Louis) or 25% (San Francisco).
Translated to the 21st century, the question is not just whether we should use NPI's for mitigation, as they are so undeniably effective, but how long you can feasibly maintain them to obtain the optimum result. One answer, if we cannot maintain NPIs for say 6 months, would have to lie in the much earlier availability of vaccines.
Part of the lesson for me is that if, in 1918, with none of this knowledge, and the NPI's were not implemented in a very systematic way nor for long enough, they could still reduce peak mortality by 50%, and overall mortality by 10-30%, how much total reduction can we achieve today with targeted approaches applied early and in an organized manner? I would suggest that a reduction from say 50% AR to 10% AR over a 12-week school closure period is certainly feasible. Which would buy us the time to ensure health services can cope, the infrastructure will hold up, and vaccines can be made as quickly as possible.