All children at infant schools will be given free school meals from 2014, Nick Clegg has announced, at a cost to the taxpayer of £600m.

On the face of it, the policy is extremely bad. Children of parents earning less than £16,190 and/or receiving income support or other kinds of welfare are already entitled to free school meals, so, other than children whose parents are unaware that they are eligible, the main beneficiaries of this policy will be the children of middle-income families.

That’s the wrong kind of redistribution. There is a decent case for helping the children of poor families who simply cannot afford to give their kids a decent packed lunch, but extending that to all children requires pretty big (and probably wrong) assumptions about parental fecklessness and state effectiveness.

But there is a complication. Trials that tested universal free school meals in schools across three local authorities between 2009 and 2011 [3] found that extending meals to some students didn’t do much, but making them universal correlated with 1.9% and 4% improvements in literacy at Key Stages 1 and 2 respectively and 2.2% and 5.5% improvements in maths at Key Stages 1 and 2. (pp 143-144)

If those numbers really were caused by making free school meals universally available (and they were more cost-effective than alternative ways of spending that money), there would seem to be a strong case for the policy. However, the authors of the government’s impact report point out that, basically, they don’t understand why this relationship exists. Neither attendance nor behaviour were affected, so they assume that free school meals led to greater classroom ‘productivity’.

What if they’re missing something and the free school meals aren’t the causal factor? Or what if nationwide implementation has bad unintended consequences we can’t foresee?

That’s exactly what happened when California rolled out a state-wide class-size reduction programme [4], an example given by Nancy Cartwright. Despite doing well in randomized controlled trials in Tennessee, in California the programme had no real effect on outcomes. The sudden need for lots of new teachers meant that more bad teachers were hired; and not all of the factors that made smaller classes helpful in Tennessee were present in California. Evidence isn’t always as transferrable as we’d like it to be.

That’s not really a good reason to think the policy will fail, but it should temper our enthusiasm for rolling it out nationwide. What works in Wolverhampton might not work in West Sussex. Indeed, what works in a particular school in Durham might not work in another school across the county. In some schools, universal free school meals might be just what the doctor ordered. In others, headmasters might think that fixing their school’s dodgy central heating would be a much better use of the money.

Ultimately, it boils down to the principle we go on about again and again here at the ASI: devolving choice down to the most basic units possible. The available evidence does suggest that “universal free school meals” are a good policy, but that evidence is quite limited. The complexity of these things means that rolling it out nationwide might not achieve what we’d hope.

Better to give schools the money and the information and let them decide what to do with it. That way, instead of basing a national policy on the results of a few randomized controlled trials, the process of experimentation and discovery can be ongoing across the country.