There was an HN thread on how even if one has statistically significant correlations, impossibility of determining causality means interventions on what one believes to be an independent variable may have no, or even the opposite, effect on what one believes to be the dependent variable. Some related link (gwern?) gave as an example the poor track record of many attempted anti-poverty interventions.
With that in mind, our approach is very KISS: rather than attempt to manipulate other things, that may or may not work, or set up entire parallel "programs", we just have the destitute buy their food and shelter on the open market, and reimburse them for the expense.