Skip to content

Are many dimensions better than one?

Over at From Poverty to Power, Duncan Greene hosted a fiery debate about how best to measure poverty, sparked by the release of the UN’s new Multidimensional Poverty Index.

The new index will complement a simpler method used in the UN Human Development Reports which relies on uniformly-weighted variables measuring life expectancy, education and income. The new method, created by researchers at the University of Oxford, combines ten different variables (including malnutrition, years of schooling, access to electricity and toilets, type of cooking fuel used, and others) and assigns them different weights.

The Oxford researchers say this is the first index covering most of the developing world to be created using micro datasets (ie household surveys), and that it is useful because it “captures a set of direct deprivations that batter a person at the same time.”

The MPI also captures distinct and broader aspects of poverty. For example, in Ethiopia 90 per cent of people are ‘MPI poor’ compared to the 39 per cent who are classified as living in ‘extreme poverty’ under income terms alone. Conversely, 89 per cent of Tanzanians are extreme income-poor, compared to 65 per cent who are MPI poor.

On Duncan’s blog, Martin Ravallion of the World Bank asks why we should add up different measures of poverty into a single index rather than getting the best data we can on individual measures, especially when weights assigned to those measures are likely to be arbitrary and controversial. (Gabriel Demombynes at the Africa Can…End Poverty blog also has a good summary of the discussion).

What is the point of creating ever more complex measures of poverty? For one, they draw attention to the importance of facets of poverty besides low income, like lack of access to education or clean water. But coming up with better measures of who is poor and how they are poor really matters if it helps allocate resources more effectively to those who need them most. It might be informative to understand why (for example) many more Ethiopians are poor under the new index than using the conventional, under-$1.25-a-day measure. But it’s hard to imagine how to find the answer without unraveling the many strands that make up the multidimensional index.

This blog frequently asks whether we should trust the figures we purport to know (for example: the malaria data cited over and over by the Gates Foundation; post-economic crisis poverty forecasts from Ravallion and colleagues; new maternal mortality figures reported in the Lancet). Aggregating different poverty measures together could also mask weaknesses in the data. Better then to measure and meet each type of deprivation separately, as best we can.

CORRECTION: In this year’s Human Development Report, the new index will be used as a complement to the existing Human Development Index, not as a replacement, as paragraph two originally stated.

This entry was posted in Data and statistics and tagged , , , , . Bookmark the permalink. Follow any comments here with the RSS feed for this post. Both comments and trackbacks are currently closed.


  1. Sam Gardner wrote:

    If poverty is multidimensional, with competing dimensions, perhaps “poverty” is not the right concept to use as a measure for development. Indeed, “income-poverty” and “power poverty” are both important issues, put it might be important to pursue them separately, taking into account the necessary trade-offs and win-wins, instead of as part of a vague mush that is called “multidimensional poverty”

    Posted August 6, 2010 at 9:01 am | Permalink
  2. John Coonrod wrote:

    I see two advantages to this measure — advocacy, and common sense. While villagers need to know multiple distinct, actionable indices (“only 1/3 of our houses have toilets!”), advocates can put a single number to good use. And that number should correlate strongly to the direct, human experience of people’s lives – which previous measures often did not.

    Posted August 6, 2010 at 9:37 am | Permalink
  3. Adam Baker wrote:

    It seems ridiculous to assign arbitrary weights to the various poverty indicators, when it would be just as possible to perform a principal component analysis, and get at the major empirical parameters of poverty. Perhaps someone has done this.

    Assigning weights willy-nilly would seem to just encode the methodological or political assumptions of the researcher into the model.

    Or just leave all the dimensions there. You’re that much more likely to creep under the .05 threshold if you have more dependent variables. :-)

    Posted August 6, 2010 at 10:54 am | Permalink
  4. Jiesheng wrote:

    The alternative otherwise is to rest back on GDP or augmented GAP figures or the HDI which I bet you also criticise. There may never be a true method in measuring how poor/rich one is , but that doesnt mean we don’t try to.

    Posted August 6, 2010 at 11:32 am | Permalink
  5. Raphael wrote:

    This raises a smaller but equally interesting and more practical question for practitioners in the field. Can we now use the “Standard of Living” subset of indicators as a good proxy for HH income / consumption? The Pearson correlation is relatively “good” at 0.8. And we all know how difficult it is to collect income data. And it’s full of error despite our best efforts! So do the 5 living standard indicators (1. access to electricity, 2. access to sanitation, 3. access to clean water, 4. cooking fuel use, 5. owning certain assets like TV/radio/etc) provide a solution to this enduring problem? Should we stop killing ourselves trying to collect income data and collect this much simpler index?

    Posted August 6, 2010 at 4:33 pm | Permalink
  6. Raphael wrote:

    Correction. The correlation is 8.9.

    Posted August 6, 2010 at 4:41 pm | Permalink
  7. Whether describing poverty or some other complex description, I would argue that any time you add granularity you accomplish the following:

    1) Avoiding the ‘error of aggregates’ is almost always beneficial for understanding causality. And a better understanding of causality tends to allow for better choice of actions.

    2) Arbitrary weights actually approximate human interpretation accurately. Humans are not absolutely quantitative, they are relatively qualitative. As such arbitrary weights, especially in this case, are often superior to any other solution.

    3) Granularity also affects politicization of the data – which is both good and bad. Simplistic numbers purporting to represent aggregates are disingenuous and misleading, as well as open to dismissal and criticism.. On the other hand, for political issues, overly simplistic analysis and simplistic representation expressed as an index is about all people can digest. So simplistic numbers are utilitarian if not accurate.

    4) None of these methods of analysis seems to account for the negative sides of development – power struggles, class warfare, factionalism, and loss of identity, nor the fact that development as we understand it, may or may not be sustainable, nor does it discuss the problem of birthrates – which is the underlying problem for all nations: too many or too little for the productivity of the group. Therefore the primary causal factor is ignored, or intentionally overlooked.

    We focus too often on aid, when the reason a nation develops a division of knowledge and labor is due as much to it’s habitual institutions, it’s belief systems, and it’s political institutions as it is to it’s geography and resources. Nor do we understand the utility of totalitarianism, republicanism, and democracy as suited to different phases of development.

    So I am often frustrated by detailed analysis of the poverty of nations, rather than a detailed analysis of what is needed to rescue a particular group of people from systemic poverty by improvement of it’s institutions. Aid is not beneficial without the broader context. It only worsens the problem.

    While people may agree upon goals, the problem for any polity, is that people cannot agree upon MEANS. This inability to agree upon means is the democratic conundrum that mandates the Iron Law Of Oligarchy. And that iron law too often leads to increase in rent-seeking (corruption and bureaucracy). Empathy is cheap. Aid is easy. But engineering a change in a society that may require we embrace values contrary to our own, and even antithetical to our own, is yet another example of our inability to agree upon means, despite the evidence of the utility of different means.

    I do not care how poor people are. I care that people are no longer poor. To eradicate poverty by external increases in production is simply maintaining the status quo. To eradicate poverty by internally empowering people is something else.

    As such, the world needs to look to China, not the west, for it’s political instruction. They have no such inhibitions.

    Posted August 7, 2010 at 12:49 pm | Permalink
  8. Brad wrote:


    Couple points. I don’t really see that principal components solves the problem. You can take the loadings on a first component that’s anchored by consumption, and those weights are sort of meaningful. But then you’re really not doing anything other than generalizing consumption, while making it impossible to interpret the units meaningfully.

    And then it’s not clear what you do with your second or third components. How do you combine those with your first component? You’re either back to arbitrary weights, or you keep them separate.

    But I don’t really see that it’s about dividing up the statistical variation into independent components. Correlated or not, how do you quantify the total crumminess of a bunch of different ways that poverty affects people’s lives?

    I’d be happy to be corrected. I think there have been a bunch of attempts, though probably not with a household-level international database.

    Posted August 8, 2010 at 12:56 pm | Permalink
  9. Scott wrote:

    Aggregating measures could mask weaknesses in the data. Unless you take time to read the study, in which it is made very clear that all the data come from a single survey for the household, rather than combining very different data sets that might have weaknesses. But hey, shoot from the hip!

    Posted August 9, 2010 at 7:45 am | Permalink
  10. Adam Baker wrote:


    I think that if you need two or three components to explain the variance (say, to the 90% level, or whatever criterion one wishes to use), then that’s an interesting fact about poverty. Assuming that all of the indicators are thought to be important, then that means that poverty is a multi-dimensional thing. I wouldn’t be surprised by that result. I think the thing to do would be to try to interpret the meanings of the different components, and carry as many of them into future analyses as possible. (This is how the Culture’s Consequences father-son team used a PCA… Hofstedter, I think is the father’s name.)

    Posted August 12, 2010 at 1:59 pm | Permalink
  11. Brad wrote:

    Thanks for the reply. You’re recommending we do a PCA, keep any components separate, and then always have, say, three dependent variables to test when we try to understand the causes of poverty and the impact of any interventions. Actually, I agree that’s a useful approach.

    But I see the question the MPI is trying to answer as about targeting: if we care about identifying those populations that are most desperately in need, how should we do that? That’s different from the “causes and impact” question you’re thinking about, but it’s still a reasonable question — even if the answer is, we have no way to know.

    Is it worse to have low income and bad health, or low income and bad education? We can’t answer that question by reporting that low income and bad health load on the same component, but education doesn’t. And maybe we can’t answer it at all. (Although if they’re all correlated, maybe we can say “don’t worry about trying to measure multidimensional poverty, you can target the least well off just by targeting the income-poor”.)

    I just think that statistical multi-dimensionality just isn’t what non-economists mean when they worry about the multi-dimensionality of poverty. Both are important, but that’s why I think PCA doesn’t do the trick for the kind of question the MPI is hoping to address.

    Posted August 12, 2010 at 6:40 pm | Permalink
  12. Adam Baker wrote:

    I read that you’re agreeing that using a PCA would be helpful for exploratory purposes. I agree that for certain purposes — such as talking to politicians — it is useful to have a single number to report. I think it’s two different applications.

    I would nevertheless be very interested to see the PCA results. Can we treat income-poverty and health-poverty as more-or-less independent problems? Thinking of the U.S. and Cuba as cases where health and wealth don’t always go together, I would certain be curious to know.

    Or, apropos to development trends, is there a single “poverty trap”? If poverty really turns out to be multidimensional, that would be an easy way to refute that simple analysis.

    Posted August 15, 2010 at 6:05 pm | Permalink

2 Trackbacks

  1. [...] This post was mentioned on Twitter by Conduit Journal, r3publican. r3publican said: Are many dimensions better than one? – Over at From Poverty to Power, Duncan Greene hosted a fiery debate about how… [...]

  2. [...] recommend that you do), please visit Gabriel Demombynes’ excellent round-up.  Also check out AidWatch, which is how I initially caught wind of the whole thing.  Finally, don’t forget to read the [...]