PDF here (via Innovators for Poverty Action, whose research is featured). After discussing the Karlan/Zinman study showing benefits for loans (which we summarize here), it continues:

Karlan is the first to warn against extrapolating too much from a single experiment. “This is the last thing in the world that I would use to develop policy,” he warns. “You’ve got to replicate.”

The trouble is that the replication just isn’t happening. For all the optimism about microfinance – and the [Karlan/Zinman] experiment only encourages that optimism – it is striking how much we do not know about when it works, and why.

This matters because non-commercial microfinance projects often depend on donor subsidies. And while microfinance has a good reputation among development professionals, that doesn’t mean guaranteed access to those subsidies. Not everyone is convinced that a donor grant is best used to subsidise a loan rather than, say, pay directly for a primary school. More credible evaluation would help preserve the programmes that deserve to be preserved.

Already, solidly held beliefs about microfinance have been shaken. The “group liability” system, in which a group of borrowers guarantee one another’s loans, is still supposed by many to be the secret behind Grameen Bank’s low default rates. But a randomised trial in the Philippines conducted by Karlan and aWorld Bank economist, Xavier Gine, found that group liability was discouraging new customers without improving repayment rates. Grameen itself quietly dropped group liability some time ago.

Another sacred cow of microfinance is that women make best use of the money – the Grameen Bank says that 97 per cent of its borrowers are women. But another randomised trial, conducted in Sri Lanka by a team of researchers including David McKenzie of the World Bank, found that male borrowers seemed to make a far higher return on their capital. As with the ZaFinCo study, it’s just one experiment in one country. Yet it raises a worrying question: for how long will donors fund microfinance projects with so little compelling evidence about exactly what kinds of project really work?

It also discusses why the Karlan/Zinman’s results might not be generalizable:

[The experiment’s] clients were ideal customers for a commercial lender: they were city-dwellers and therefore cheap to reach; they were poor enough to want loans but rich enough that the loans were profitably chunky. A peasant farmer in Ethiopia or Sudan ticks none of those boxes: living in the middle of nowhere, he is expensive to reach and he is so poor that he can only afford tiny loans. Trapped in a barren economic ecosystem, he has no job that a ZaFinCo-style loan can help preserve, and no business prospects either. A mere loan will not catapult him into the ranks of the entrepreneurial class. Then there are the destitute, the disabled, the elderly and the orphans. Such people cannot repay loans at a rate that would cover costs. Heavy subsidies or outright grants would be needed. “All people are entrepreneurs,” says Muhammad Yunus. If only he were right.

And indeed, a more recent randomized trial of microfinance has preliminarily found far less encouraging results (discussed in this interview on Philanthropy Action, co-maintained by GiveWell Board member Tim Ogden).