Vaccine development as a decision problem

This post by Alex Tabarrok hits all the right notes:

At current rates, the US economy is losing about $40 billion a week. Thus, if $20 billion could advance a vaccine by just one week that would be a good deal. . . . It might seem expensive to invest in capacity for a vaccine that is never approved, but it’s even more expensive to delay a vaccine that could end the pandemic.

I [Tabarrok] am also concerned that OWS is narrowing down the list of candidates too early . . . These are all good programs and one of them will probably be successful but we also want to support some long-shots because a small probability of a very big gain is still a big gain. . . .

The Accelerating Health Technologies team that I am a part of collected data on over 100 vaccine candidates and their characteristics. We then created a model to compute an optimal portfolio. We estimated that it’s necessary to have 15-20 candidates in the portfolio to get to a 80-90% chance of at least one success and that you want diverse candidates because the second candidate from the same platform probably fails if the first candidate from that platform fails. Moderna and Pfrizer are both mRNA vaccines–a platform that has never been used before–while AstraZeneca, Johnson and Johnson and Merck are using somewhat different viral vector platforms . . .

One way to diversify the portfolio is to make deals with other countries to avoid the prisoner’s dilemma of vaccine portfolios. The prisoner’s dilemma is that each country has an incentive to invest in the vaccine most likely to succeed but if every country does this the world has put all its eggs in one basket. To avoid that, you need some global coordination. . . .

I have not looked at the details here at all, but I like the general approach. I like the awareness of uncertainty, variation, costs, benefits, contingencies, and politics. We https://statmodeling.stat.columbia.edu/2020/05/03/curing-coronavirus-isnt-a-job-for-social-scientists/“>talked about how social science can used to fight coronavirus. This sort of thing is an example.

I hope this all can be done in conjunction with Bayesian inference for vaccine effects in the general population and sub-populations. I’d hate to see all this careful design and decision analysis stapled onto crude deterministic inference procedures based on statistical significance.