What are epidemiological models for? You can use models to inform policy and other decision-making. But you can’t use them to manufacture a number that you can advertise in order to draw praise. That’s what the government’s excuse appears to be vis-à-vis the number of deaths averted by India’s nationwide lockdown.
When the government says 37,000 deaths were averted, how can we know if this figure was right or wrong? A bunch of scientists complained that the model wasn’t transparent, so its output had to be taken with a cupful of salt. But as an article published in The Wire yesterday noted, these scientists were asking the wrong questions – that the number of deaths averted is only a decoy.
So say the model had been completely transparent. I don’t see why we should still care about the number of deaths averted. First, such a model is trying to determine the consequences of an action that was not performed, i.e. the number of people who might have died had the lockdown not been imposed.
This scenario is reminiscent of a trope in many time-travel stories. If you went back in time and caused someone to do Y instead of X, would your reality change or stay the same considering it’s in the consequent future of Y instead of X? Or as Ronald Bailey wrote in Reason, “If people change their behaviour in response to new information unrelated to … anti-contagion policies, this could reduce infection growth rates as well, thus causing the researchers to overstate the effectiveness of anti-contagion policies.”
Second, a model to estimate the number of deaths averted by the lockdown will in effect attempt to isolate a vanishingly narrow strip of the lockdown’s consequences to cheer about. This would be nothing but extreme cherry-picking.
A lockdown has many effects, including movement restrictions, stay-at-home orders, disrupted supply of essential goods, closing of businesses, etc. Most, if not all, of them are bound to exact a toll on one’s health. So the number of deaths the lockdown averted should be ‘adjusted’ against, say, the number of people who couldn’t get life-saving surgeries, the number of migrant labourers who died of heat exhaustion, the number of TB patients who developed MDR-TB because they couldn’t get their medicines on time, even the number of daily-wage earners’ children who died of hunger because their parents had no income.
So the only models that can hope to estimate a meaningful number of deaths averted by the lockdown will also have simplified the context so much that the mathematical form of the lockdown will be shorn of all practical applicability or relevance – a quantitative catch-22.
Third, the virtue of the number of deaths averted is a foregone conclusion. That is, whatever its value is, it can only be a good thing. So as an indisputable – and therefore unfalsifiable – entity, there is nothing to be gained or lost by interrogating it, except perhaps to elicit a clearer view of the model’s innards (if possible, and only relative to the outputs of other models).
Finally, the lockdown will by design avert some deaths – i.e. D > 0 – but D being greater than zero wouldn’t mean the lockdown was a success as much D‘s value, whatever it is, being a self-fulfilling prophecy. And since no one knows what the value of D is or what it ought to be, even less what it could have been, a model can at best come up with a way to estimate D – but not claim a victory of any kind.
So it would seem the ‘number of deaths averted’ metric is a ploy disguised as a legitimate mathematical problem whose real purpose is to lure the ‘quants’ towards something they think challenges their abilities without realising they’re also being lured away from the more important question they should be asking: why solve this problem at all?