Tuesday, November 18, 2008

These things do happen...

You have carefully calculated your sample size based on some criteria (power calculation, impact size, budget, etc.). You have randomly assigned your samples into treatment and control groups. Your counterpart, in this case District governments, have agreed to commit to the random assignment.

Then, one morning, your local government contact called you:
How are you? Been a while since the last time we talked. By the way, sorry to inform you, but we have decided to give treatment to village X. I know, initially it was a control village. But we thought that the village is in a desperate need to get treatment. So, in fact, we have already started treating village X.
And this was not the first time you received that kind of surprise phone call. And out there, many still believe that a big donor like the Bank is a kind of super institution that is able to dictate governments here and there to get what it wants.

Note: anyway, in randomized trial, there are ways to deal with this; you can choose to apply the Intention to Treat analysis, or the Treatment of the Treated.

4 comments:

  1. ape, i've been wondering whether randomised experiments in economics are really random -- because in many experiments, the control group know their status (say, a village who know that the neighbouring village are getting something they don't, or the lottery losers who know they are losing while others are winning), and so might behave differently.

    as this is different to the kind of randomised experiments we have in psychology and medicine, where the control group simply have no idea of their status, has any economist looked closely at this issue?

    ReplyDelete
  2. Tirta,
    they are called randomized experiments for the random assignment of treatment and control.

    the fact that respondents might behave differently than they would without the assignment is another issue. a lot of literature has dealt with this, as I think you are familiar with the Hawthorne effect, placebo or pygmalion effect, etc. these are problems that would also arise in psy or med-type randomized trials, aren't they?

    and, in many case, econ randomized experiments (or 'social experiment') look at the 'intention to treat.' meaning, yes, we are aware of such spillover, attrocity or behavioral problems. but that's what would happen in the real life anyway (i.e. when the govt imposes some intervention).l

    ReplyDelete
  3. yes, 'randomised' means random assignment of treatment and control -- which necessitates subjects NOT knowing which condition they're in. this is why you can be absolutely sure that the only differing variable between the two groups is *treatment* itself, not *treatment-and-knowing-that-we-are-being-treated*.

    such problems rarely exist for psych and med experiments, because one can set-up another condition to *control* for the placebo (for testing headache pill, for instance, you would have one group given real pills, another group without any pill, and yet another group with fake pills -- all naive about the existence of the other groups).

    i know it's awfully hard to translate the level of lab control we in psych and med have to the real field where economists operate. and so i wonder if, for instance, the 'intention to treat' approach is widepsread enough, given that i see many, if not most, econ experiments as subject to the problem of knowledge and expectation.

    ReplyDelete
  4. Tirta,
    believe me in many RE cases, subjects don't know if they are part of the experiment. yes, they may know they are receiving something (iron pills, anti-worm medicine), but not they don't necessarily know that they are part of an experiment.

    in some cases, subjects know that they do or don't get something, compared to the other. village receiving projects, for instance there are ways to do this. for example, rotating treatment/control or phase-in/phase-out method.

    in other cases, Hawthorne or placebo effects do create problems. just like in psy/med experiments. but you're right, econ/social RE are much harder to control. my problems with the local governments who like to tamper the random assignment is just a so common.

    but i'll tell you what -- the intention to treat is almost a standard approach for randomistas. it is used more than the treatment of the treated.

    ReplyDelete