Mead Over at the CGD discusses a recent debate between Martin Ravallion, frequent critic of the multidimensional poverty index (MPI), and James Foster, one of its inventors. For those of you not familiar with the MPI, you can read about it here, where I briefly explain and discuss the measure.
Ravallion criticises the MPI for using potentially arbitrary weights to combine several different measures of poverty into single, hard to interpret index. When researchers start assigning weights to create composite indices, they are implicitly making judgements on how we value different measures of poverty. For instance, if our MPI of interest is constructed using 2 x(access to water) + 1 x (asset poverty) (gross simplification), we are implicitly saying we care about access to water twice as much as we do about asset poverty.
This had led Ravallion (and others) to suggest that governments consider each measure separately, rather than compiling them into something that lacks an immediate interpretation:
Ravallion asked the audience to remember their last appointment with a physician for a general physical examination â€“ a â€śroutine checkup.â€ť He asked, â€śWould you really want the doctor to summarize all the information on all the tests with a single composite index of you health?â€ť
Foster objects, noting that battlefield medics will collapse all these measures into one index (i.e who is closest to death?) in order to establish priority. Mead Over manages to get to the heart of the disagreement: Ravallionâ€™s approach makes no assumption of the MPI-userâ€™s value function (i.e., what they care about most), where Fosterâ€™s assumes that someone is going to be imposing some sort of a value/objective function, in which case weighting and merging into an index makes sense.
If we wanted to establish a value function for weighting the MPI, what might it look like? Over suggests using some empirics to help us determine which measures of poverty are to be given higher priority (i.e. a higher weight in the MPI).
With household level panel data long enough to follow a child from birth to maturity, one could use a familyâ€™s early scores on all the dimensions of the MPI in order to explain the subsequent escape from poverty of that familyâ€™s grown children. For the country context in which the study is performed (and subject to the usual conditions for a regression to estimate causal pathways), the multiple regression coefficients would provide weights for an MPI which has the specific meaning that Foster suggests would be useful.
This is an idea worth exploring, if anything because I havenâ€™t thought about this approach before.
It strikes me that this would be an incredibly difficult exercise if we were determined to discover the causal connection between the myriad of MPI poverty indices and subsequent exit from poverty. Most of these measures will be subject to a high degree of endogeneity, and finding a reliable set of instrument for the k measures we are worried about seems unlikely. Even if the coefficients on such a regression can be interpreted causally, the routes of escaping poverty might have changed across time: for example, access to roads may have been the secret to escaping poverty 10 years ago, but mobile phones may now be the key to escaping 10 years from now, yet our regression would tell us to focus on getting people access to better roads.
Assuming the casual method is impractical or undesirable, the coefficients from a plain-vanilla regression might still tell us something about what kinds of households we should be targeting with social assistance. For instance, if we find that access to clean water is more highly predictive of subsequent access, we might want to aim our interventions at people with poor access to clean water. When all this is added up, the regression-weighted MPI tells us how well weâ€™re likely to see a future reduction in poverty (as long as generation effects donâ€™t get in the way here as well).
There are still problems though. More often than not, multiple non-income poverty measures tend to be heavily correlated, (many economists believe that they are all essentially measuring the same thing) so throwing those all into a regression together might not tell us as much as weâ€™d like. It is also unclear what one would use for our measure of future exit from poverty (the dependent variable in the regression). We canâ€™t use future MPI, because we havenâ€™t decided on the weights yet. The problem here is that whatever measure we choose, we implicitly declare this to be our preferred measure of poverty. If we just use income poverty, then itâ€™s possible that we are not going to find a better predictor of future exit from income poverty than current income poverty, at which point we should just throw in the towel.
Hereâ€™s an idea: why donâ€™t we let the poor set the objective function? As Iâ€™ve written about before, there are several examples of successful preference aggregation, such as the US armyâ€™s use of random sampling and ranking to create the points system used for discharging soldiers at the end of WWII. Governments could randomly sample the population, asking them to rank different people with different `bundlesâ€™ of multidimensional poverty (i.e. who is worse off, someone with no access to water, or someone living on less than $1 a day?), and then could aggregate the responses to generate weights for the MPI. Such an index then would accurately reflect the aggregate desires and aspirations of the poor, rather than the â€śexpert-weightedâ€ť measures that Mead Over is sceptical of.