Sunday, August 1, 2010

Ravallion and Alkire on Multidimensional Poverty Index (MPI)

Martin Ravallion on Multidimensional Poverty Index (MPI)-- via Duncan Green's blog. For introductory reading on MPI see this blog entry.

“Everyone agrees that poverty is not just about low consumption of market commodities by a household.  There are also important non-market goods, such as access to public services, and there are issues of distribution within the household. It is agreed that consumption or income poverty measures need to be supplemented by other measures to get a complete picture.

But does that mean we should add up the multiple dimensions of poverty into a single composite index? Or should we instead measure consumption poverty with the best data available, while also looking for the best data on other dimensions of poverty as appropriate to the country context?

The Oxford Poverty and Human Development Initiative (OPHI) has recently launched a Multidimensional Poverty Index (MPI), and calculated it for over 100 countries.   The MPI is a composite of indicators selected for consistency with the UNDP’s famous Human Development Index (HDI). The HDI uses aggregate country-level data, while the MPI uses household-level data, which is then aggregated to country level. The index has ten components; two represent health (malnutrition, and child mortality), two are educational achievements (years of schooling and school enrolment), and six aim to capture “living standards” (including both access to services and proxies for household wealth).  The three broad categories–health, education, and living standards–are weighted equally (one-third each) to form the composite index.  

One can debate the precise indicators chosen for the MPI by the Oxford team (who are clearly aware of the many data concerns). For example, the MPI’s six “living standard” indicators are likely to be correlated with consumption or income, but they are unlikely to be very responsive to economic fluctuations.  The MPI would probably not capture well the impacts on poor people of economic downturns (such as the Global Financial Crisis) or rapid upswings in macro-economic performance.

The precise indicators used in the MPI were not in fact chosen because they are the best available data on each dimension of poverty. Rather they were chosen because the methodology used by the MPI requires that the analyst has all the indicators for exactly the same sampled household. So they must all come from one survey. There is much better data available on virtually all of the components of the MPI, but these better data can’t be used in the MPI since they are only available from different surveys. This aspect of their methodology greatly constrains the exercise. If one chooses not to form the composite at household level but to look instead at the separate dimensions of poverty one is free to choose the best available data on each dimension of poverty.

There is a deeper concern about the MPI, which holds even if the best data all came from just one survey. The index is essentially adding up “apples and oranges” without knowing their relative price. When one measures aggregate consumption from household-survey data for the purpose of measuring poverty, as in the World Bank’s “$1 a day” measures, one relies on economic theory, which says that (under certain conditions) market prices provide the correct weights for aggregation. We have no such theory for an index like the MPI. A decision has to be taken, and no consensus exists on how the multiple dimensions should be weighted to form the composite index.

On closer scrutiny, the embedded trade-offs (stemming from the weights chosen by the analyst) can be questioned, and may be unacceptable to many people.  In the context of the HDI, I pointed out 15 years ago that by aggregating GDP per capita with life expectancy the HDI implicitly put a value on an extra year of life, and I showed that this value rises from a very low level in poor countries to a remarkably high level in rich ones (4-5 times GDP per capita).   If it was made clearer to users, I expect that they would question this trade-off embedded in the HDI.

The MPI index faces the same problem. How can one contend (as the MPI does implicitly) that the death of a child is equivalent to having a dirt floor, cooking with wood, and not having a radio, TV, telephone, bike or car?  Or that attaining these material conditions is equivalent to an extra year of schooling (such that someone has at least 5 years) or to not having any malnourished family member?  These are highly questionable value judgments. Sometimes such judgments are needed in policy making at country level, but we would not want to have them buried in some aggregate index.  Rather, they should be brought out explicitly in the specific country and policy context, which will determine what trade off is considered appropriate; any given dimension of poverty will have higher priority in some countries and for some policy problems than others.

Poverty is indeed multidimensional.  But it is not obvious how a composite multidimensional poverty index such as the MPI contributes to better thinking about poverty, or better policies for fighting poverty.  Being multidimensional about poverty is not about adding up fundamentally different things in arbitrary ways. Rather it is about explicitly recognizing that there are important aspects of welfare that cannot be captured in a single index.”

Sabina Alkire, the one who came up with MPI, responded:

“As Martin Ravallion points out, we agree that poverty is multidimensional. The question is whether our efforts to incorporate multiple dimensions into the very definition of who is poor and the measurement of poverty “contributes to better thinking about poverty, and to better policies for fighting poverty.” Let me explain what I think the MPI adds.

The MPI measure has meaning in itself and can also be broken down immediately into its component parts.  Every time you see an MPI figure – for a person, an ethnic group, a state, or a country – you know that it also contains what could be thought of as a drop-down menu in two layers. The first layer shows incidence and intensity. The second breaks the MPI down by indicator and shows what poverty is made of.

If we know someone is income poor, we do not know if they are also illiterate or malnourished. If we know someone is multidimensionally poor, we can unpack the MPI to see how they are poor. That is one added value of our methodology.  That is why we call it a high resolution lens: you can zoom in and see more.

This feature could add value to the MDG indicators too. These show us the percentage of people who are malnourished, and the rate of child mortality and many other things, but not how the deprivations overlap. If 30% of people are malnourished and 30% of children are out of school, it would be useful to know if these deprivations affect the same families or different ones. With the MDG indicators we cannot see that; with the MPI, we can. Of course not for all MDG indicators, but it’s a start.

For example, the Somali have the highest multidimensional poverty of all ethnic groups in Kenya followed by the Masai. Looking at the MPI drop-down menu, we see that 96% of Masai are poor and 88% of the Somali. But poverty among the Somali is more intense: on average they are deprived in 67% of dimensions; the Masai in 62%. Zooming in further we note that the Somali are more deprived in education and child mortality, whereas malnutrition and standard of living indicators are worse among the Masai. So the MPI opens out into a wider field of information.

The other thing the MPI does is clean data of anomalies and focus on poor people. While indicators drawn from different surveys are tremendously useful for many purposes, they do not identify who is multidimensionally poor, so every MPI poor person experiences multiple deprivations. Consider a self-made millionaire who didn’t go to school. A MDG indicator includes this millionaire in the percentage of people who are uneducated. The MPI does not – if she’s not deprived in anything else, she’s not considered poor. In times of tighter fiscal resources we focus on people who are deprived in several things at the same time.

So, the MPI – and the general methodology it uses that James Foster and I developed – adds value because of how it evaluates poverty. The method first determines the dimensions in which a person is deprived, and then ‘adds up’ that person’s deprivations using weights that reflect the relative importance of each deprivation. A person who is sufficiently ‘multiply deprived’ is considered poor. We measure multidimensional poverty as the incidence (or the percentage of the population that is poor) times the intensity (or the average percentage of deprivations poor people experience). Unlike the HDI, this construction does not add up achievement levels, which requires strong assumptions concerning the variables in question as Martin noted. Instead, we add up deprivations, which does not.

OK, now to the issue of weights. Income poverty aggregates within a country using actual or imputed prices (these are critical for fixing the income poverty standard across countries and time). Setting prices is not unproblematic in practice, particularly in Colombia where I am writing from. Indeed the Presidential address to the 2010 American Economic Association raised concerns such as the prices attributed to housing (Deaton 2010). Chen and Ravallion 2008 carefully review the robustness of their results to different pricing approaches.

As Martin observed, instead of using prices, the MPI sets weights as value judgements. Amartya Sen among others sees this feature as a strength not an embarrassment: “There is indeed great merit… in having public discussions on the kind of weights that may be used” (1997a).

In extensive writings, Sen presents several pertinent observations in favor of setting weights: first, prices may not exist for some aspects of poverty (morbidity, mortality, illiteracy) but giving zero weight to these does not seem right either. Second, setting precise weights may not be necessary: comparisons may be robust to a range of weights. Third, the weights trigger public debates which may be constructive as policy makers weight tradeoffs in practice anyway.

The key, Sen suggests, is to make the weights explicit: “It is not so much a question of holding a referendum on the values to be used, but the need to make sure that the weights – or ranges of weights – used remain open to criticism and chastisement, and nevertheless enjoy reasonable public acceptance” (1997b; see also Decancq and Lugo 2008).

Given this situation, Maria Emma Santos and I proceeded in a very practical way in the MPI. First, the weights are not buried; they are totally transparent (1/3 per dimension, and each indicator within a dimension equally weighted). If people disagree with these weights, they can propose improvements and also recalculate with different weights to check robustness. Second, the weights give some non-zero value to each dimension, which is a starting point. Third, the MPI fixes weights between countries to enable cross-national comparisons; alongside this we strongly encourage countries to develop national measures having richer dimensions, and indicators and weights that reflect their context as Mexico did and Colombia is doing. Fourth, we weight the three dimensions equally, this was corroborated by expert opinion (Chowdhury and Squire 2006), helps make it easy to understand (Atkinson et al. 2002) and at least for the HDI is quite robust (Foster McGillivray and Seth 2009). We do need to create robustness tests for MPI weights, and new methodologies of analysis to guide policy, and OPHI plan to work with other researchers on these. But the key thing is that the present MPI weights are transparent, and critical scrutiny of them is welcomed.

Finally, both previous blogs mentioned data constraints. Duncan criticised the MPI for including only three dimensions, “partly because it still relies on existing data sets.” Well, actually data constraints are the only reason only these three dimensions appear. We and the HDRO wish to include others without losing focus: indeed OPHI’s other research theme highlights the need to gather internationally comparable data on ‘Missing Dimensions’ of poverty – violence, informal work, disempowerment, and isolation/humiliation — given the importance these have in poor people’s lives. Our methodology is flexible enough to accommodate additional dimensions as they become available and we are eager to do so.

Finally, as Martin observed, our data must come from the same survey or from matched surveys. Yet multi-topic surveys have expanded rapidly, especially since the MDGs. The MPI is not perfect, but it uses these surveys to explore joint distribution – the deprivations that batter poor people’s lives at the same time. Such a multidimensional poverty measure complements existing tools. So though no measure is enough, we hope this work will enable others fight poverty and empower poor people more effectively.”