Simpson's paradox and Social Impact Bonds.

 Social Impact Bonds have gotten a bad press recently. Currently, they are not tradable and operate as contingent contract- i.e. the Govt. pays out if a target- like reduced recidivism- is met, otherwise the contractor is out of pocket.

Clearly, a bit of Agency Capture can make SIBs profitable.  Targets are manipulable, you can reduce recidvism by training delinquents to be better criminals or hire them better lawyers- but, hopefully, only reputable non profits, unlike Seedco, would be granted such contracts in the first place. As such, assuming good faith and common knowledge, there seems no a priori reason why Contractors should not be paid on a contingency basis for any desired Social Impact. Exogenous factors which wipe out profits might just as well enhance them and, in any case, could be insured against.

However, there are two different amalgamation problems which undercut this view. The first concerns the power law such that a small proportion of a given population produces the bulk of the negative social impact being targeted. This will militate for different treatment, unless this is specifically forbidden (in which case the Govt. may as well perform the service) causing the population to split giving rise to Yule-Simpson type effects which muddy the water re. potential scalability.
The second amalgamation problem has to do with who is currently bearing the risk of Social impact failure and how they are off setting that risk. Once again, a power law arises such that a small proportion of the population bears a higher share of risk and how they off set that risk has to be understood before commissioning a SIB. At the margin, risk may be being offset by median voter behavior or some other effective 'Voice' strategy vital to political health. This may not be scalable- for example, from the District to the State level. Here, again, a Yule-Simpson effect arises such that the success of a pilot SIB scheme provides no guidance as to its scalability.

By contrast, performance of unilateral, not contingent, contracts, don't face such aggregation problems and are eminently scalable, benefiting as they would from reduced uncertainty because of the Law of large numbers.

A more profound objection to this approach has to do with the great plasticity of social, as opposed to individual, morbidity such that Rossi's metallic laws hold true. After all, it's only in Lake Woebegone that all the kids are above average.