The Banerjee and Newman (1993) theory model of poverty traps is very elegant, and has been influential for both development research and policy. The basic idea of this type of model is that there are only two occupations someone can do: a subsistence activity that requires no fixed costs, and a more profitable activity that requires a lumpy capital investment. With borrowing constraints, in this model, this results in multiple equilibria. There are two stable steady state levels of assets: start below some level and your returns are very low and you converge to the subsistence level, start above this level, and you have high returns and can converge to the higher level. This gives rise to a classic S-shape dynamic when plotting assets in the future against assets today.
In a QJE paper with the grandiose title “Why do people stay poor?”, Balboni et al. (2022) argue that the answer to this title question is because poor people lack enough assets to go into this more profitable occupation. The people here are poor women (average age 35) in poor villages in rural Bangladesh, and the asset is a single cow (valued at around $490 on average). They have a randomized experiment of a graduation program that gives these women a cow, as well as some consumption support and training. They have an experiment with a large sample (over 6,000 extremely poor women) which they track over a long period of time (focusing mainly on a 4-year follow-up, but also looking over 11 years). The main result is seen in this already-famous figure, which shows the classic S-shape of this theory model, where an S-shaped curve crosses the 45 degree line from below. They estimate that there is a threshold at $504, just $14-16 above the average asset value they gave: households above this level after getting their cow see their assets grow, while those who were below this threshold see their assets shrink. They argue based on this that large asset transfers are “an effective means of getting people out of poverty traps and reducing global poverty”.
Balboni et al. (2022) figure showing S-shaped asset dynamics
Aart Kraay and I wrote a review paper for the JEP in 2014 assessing the evidence for whether various types of poverty traps exist. We noted there that the closest analogs we had found to this Banerjee and Newman type case were work by Chris Barrett and co-authors looking at rural pastoralists in Ethiopia and Kenya, where in very remote locations with only cows as production technology and no access to credit, these dynamics might apply. At the time, I saw this Balboni et al. (2022) paper as another example of this, with better identification. I had two main objections at the time to this being the answer to their title of “why do people stay poor?”
· First, note that even the upper steady state in the Balboni et al. figure is at around $800, not so far from the lower steady-state of $410. So taking their result literally, these are all poor people, at most it was explaining why some poor people were slightly less poor than others. See Lant Pritchett for a classic “Lant-rant” on how setting poverty lines very low is one explanation for why giving people chickens or cows is seen as the solution to poverty.
· More importantly, the bigger question is why is this the only production technology? Especially as soon as one moves to less remote places, there are a whole range of productive activities, each with different asset requirements – and so the production function convexifies very quickly. Balboni et al. note that their sample are women in their mid-30s, with only 0.49 years of education, who are unable to migrate. They write down and estimate a model of misallocation, and say that in the absence of credit constraints, only 2% of households would be best specializing in wage labor, and only 1% do livestock rearing, whereas optimally 90% should be raising cows. But I would argue the much bigger misallocation and reason they are poor is that these women have no education, and are living in the wrong place. In a long-term follow-up of an ultra-poor program in India, Banerjee et al. (2021) find that one of the main channels for sustained impacts there is that “is that treated households take better advantage of opportunities to diversify into more lucrative wage employment, especially through migration.”
Why I am discussing this now?
This paper is getting a lot of renewed attention because of new working paper by Dean Karlan, Amol Singh Raswan and Christopher Udry, which also has a noteworthy title “The Sisyphean pursuit of evidence for poverty traps”. This paper looks at how robust this evidence for S-shaped asset dynamics is.
1. These asset dynamics do not show up in 7 other graduation experiments: they estimate similar models with data from graduation experiments in Ethiopia, Ghana, Honduras, India, Pakistan, Peru, and another graduation program in Bangladesh, and don’t see this S-shape in any of these cases.
2. They identify two important issues that suggest the S-shape in Balboni et al. is not robust to functional form issues around log (x/a +1) transforms, and may confound geographic differences with wealth differences.
At the four-year follow-up, 8.4% of households reported zero for all the asset questions. To avoid dropping these households while looking at asset growth, Balboni et al. use the transform log(assets/1000 + 1). That is, expressing assets in thousands of BDT, and adding 1. Recent work since this paper was written, such as Chen and Roth, have emphasized how sensitive results can be to the exact way zeros are handled. Karlan et al. consider four alternative specifications (log(assets/10,000 + 1), log(assets/100 + 1), log (assets + 1), and assets in levels. Their figure below shows they only get an estimated S-shaped function that crosses the 45 degree line when using log(assets/100 +1), but none of these alternatives. In particular, look at the curve in levels: whatever level households started with, they have much higher levels on average 4-years later.
The S-shaped curve is not robust to functional forms around log(x/a+1).
Secondly, note that Balboni et al. did not randomize different asset amounts to different households. So variation in baseline wealth + the cow is not exogenous, only pre-determined. Karlan et al. note that while the program gave almost every treated person a cow, the cow is valued based on the price of a cow in the village each household lives in. The result is that households that look wealthier to start with are much more likely to be from villages with high cow prices. They find that even in the control group, ultra-poor households living in areas with high cow prices had their assets grow faster than those living in areas with low cow prices. So what looks like differences in asset growth due to wealth differences, could be just picking up that some locations are growing faster than others. The places with high cow prices are in areas with less river erosion, more NGO activity, greater access to electricity, more educated households, and many other differences. So this is a nice illustration of the point that when we look at treatment heterogeneity by some baseline characteristic, that baseline characteristic might be picking up a lot more than we think it is.
So is the search for poverty traps a Sisyphean pursuit?
Karlan et al. argue that they do not interpret their re-analysis as evidence against asset-threshold poverty traps, but rather an illustration of how hard it can be to detect them. In a reply, Balboni et al. also stress this difficulty once one starts allowing for heterogeneity, so that any threshold could differ for each household. I agree that taking this theory literally where there are really only two occupations is unlikely to be appropriate for explaining why many people remain poor. But as we discuss in Kraay and McKenzie, the idea of a single big discrete jump in production technologies may make more sense when we think of geography: whether you are in a poor rural Bangladeshi village versus a city in Bangladesh, versus migrate to another richer country can make a huge difference to poverty. So maybe rather than looking for cows, it should be buses and plane tickets…
Join the Conversation