At the end of 2017, we named Evidence Action’s No Lean Season one of GiveWell’s nine top charities. Now, GiveWell and Evidence Action agree that No Lean Season should not be a GiveWell top charity this year, and Evidence Action is not seeking additional funding to support No Lean Season’s work at this time.
This post will discuss this decision in detail. In brief, we updated our assessment of No Lean Season, a program that provides loans to support seasonal migration, based on preliminary results Evidence Action began discussing with us in July from a study of the 2017 implementation of the program (hereinafter referred to as “2017 RCT”). These results suggested the program, as implemented in 2017, did not successfully induce migration. Taking this new information into account alongside previous studies of the program, we and Evidence Action do not believe No Lean Season meets our top charity criteria at this time.
Evidence Action’s post on this decision is here.
GiveWell’s mission is to identify and recommend charities that can most effectively use additional donations. While it may be disappointing for a top charity to be removed from our list of recommendations, we believe that adding and removing top charities from our list is an important part of our process. If our top charities list never changed, we would guess we were (a) acting too conservatively (i.e. not being open enough to adding new top charities), or (b) not being critical enough of groups once they’ve been added to our list (i.e. not being open enough to removing existing top charities).
We believe this decision speaks positively of Evidence Action and demonstrates our mutual commitment to updating our views based on new evidence. GiveWell has interacted with hundreds of organizations in our history, and very few have subjected their programs to a rigorous study in the way that Evidence Action did last year and, at smaller scale, in 2014. We’re excited to work with a group like Evidence Action that is committed to rigorous study and openness about results.
Summary
In this post, we will discuss:
- The history of GiveWell and No Lean Season. (More)
- How the 2017 RCT updated our views of No Lean Season. (More)
- Conclusion
GiveWell and No Lean Season
No Lean Season provides support for low-income agricultural workers in rural Bangladesh during the time of seasonal income and food insecurity (“lean season”). The program provides small, interest-free loans to support workers’ temporary migration to seek employment. No Lean Season is implemented by RDRS Bangladesh; Evidence Action provides strategic direction, conducts program monitoring, and provides technical assistance, among other functions. Evidence Action developed No Lean Season as part of its Beta portfolio, which is focused on prototyping and scaling cost-effective programs.
GiveWell began engaging with No Lean Season as a potential top charity in 2013, when we began to explore making an Incubation Grant to support its scale-up. We saw No Lean Season as a promising program that lacked the track record to be considered for a top charity recommendation at that time. We describe our initial interest in the program in a February 2017 blog post:
We approached Evidence Action in late 2013 to express our interest in supporting the creation of new GiveWell top charities.
In March 2014, Good Ventures made a $250,000 grant to Evidence Action to support the investigation and scale-up of promising programs.
Since then, Good Ventures has made three additional grants totaling approximately $2.7 million to support the program’s scale-up.
No Lean Season continued to test and scale their program with this and other support. We decided to recommend No Lean Season as a top charity in late 2017. We based our recommendation on three randomized controlled trials (RCTs) of the program. (We generally consider RCTs to be one of the strongest types of evidence available; you can read more about why we rely on RCTs here.)
Two of the RCTs (conducted in 2008 and 2014) indicated increased migration, income, and consumption for program participants. In the third RCT, which was conducted in 2013 and has not been published, the program is considered to have failed to induce migration, potentially due to political violence that year. We discuss the RCT evidence in greater depth in our intervention report on conditional subsidies for seasonal labor migration in northern Bangladesh.
Weighing the evidence, the cost of the program, and the potential impacts, we decided No Lean Season met our criteria to be named a top charity in November 2017. We summarized our reasoning in our blog post announcing our 2017 list of top charities, and noted the risks of this recommendation:
Several randomized controlled trials (RCTs) of subsidies to increase migration provide moderately strong evidence that such an intervention increases household income and consumption during the lean season. An additional RCT is ongoing. We estimate that No Lean Season is roughly five times as cost-effective as cash transfers (see our cost-effectiveness analysis).
Evidence Action has shared some details of its plans for monitoring No Lean Season in the future, but, as many of these plans have not been fully implemented, we have seen limited results. Therefore, there is some uncertainty as to whether No Lean Season will produce the data required to give us confidence that loans are appropriately targeted and reach their intended recipients in full; that recipients are not pressured into accepting loans; and that participants successfully migrate, find work, and are not exposed to major physical and other risks while migrating.
As indicated above, No Lean Season conducted an additional RCT to evaluate its program during the 2017 lean season (approximately September to December), the preliminary results of which indicate the program failed to induce migration. With the evidence from the 2017 RCT, the case for the program’s impact and cost-effectiveness looks weaker.
Our updated perspective on No Lean Season
The 2017 RCT was a key factor in the decision to remove No Lean Season from our top charities list. Below, we discuss:
- What the 2017 RCT found.
- Our interpretation of the RCT results.
- No Lean Season’s future.
What did the 2017 RCT find?
The 2017 RCT was a collaboration between Evidence Action, Innovations for Poverty Action, and researchers from Yale University, the London School of Economics, and the University of California, Davis. In a preliminary analysis shared with GiveWell in September 2018, the researchers did not find evidence for a negative or positive impact on migration, and found no statistically significant impact on income and consumption.[1]
However, the implementation of the program during the 2017[2] lean season and the evaluation of it differed from previous iterations. No Lean Season operated at a larger scale in the fall of 2017 than it had previously, offering loans to 158,155 households, compared with 16,268 households in 2016. Relative to earlier versions of the program, the program in 2017 involved (a) higher-intensity delivery of the intervention (offering loans to most eligible individuals) and (b) broader eligibility requirements (the eligibility rate in 2017 was 77 percent, compared with 49 percent in 2016).[3]
At this point, neither GiveWell, nor No Lean Season, nor the researchers feel we have a conclusive understanding of why the program failed to induce migration. However, No Lean Season and the researchers are exploring various hypotheses about what may explain the failure to induce migration, and they note that some suggestive evidence supports some hypotheses more than others. The researchers have posited several possibilities:
- The way the program was targeted in 2017 was suboptimal. The Migration Organizers, who survey households for eligibility and offer and disburse loans (more detail here under “Migration Organizers”), may have focused their efforts on the individuals that were seen as most likely to migrate, rather than those who needed a loan to afford migration. The use of loan targets during implementation may have inadvertently incentivized this behavior.[4] If, for example, loan officers mostly made loans to people who would have migrated regardless of receiving a loan, this could have led to the lack of impact on migration found in the study.
- The 2017 lean season was particularly bad for the program. The researchers note that severe flooding and associated implementation delays in some regions may have caused problems in 2017. The researchers plan to look more closely at the regions that experienced flooding, though they note that they don’t have the data necessary to make experimental comparisons.[5] In addition, a 2013 trial may have failed due to issues that were specific to the year of that trial, such as increased labor strikes.
- There exists another (currently unknown) reason why this program won’t work at scale. Conditions in Bangladesh may have changed, negative spillovers (harmful impacts for individuals who did not receive loans) may cancel out gains, or pilot villages may have been strategically picked in earlier trials.[6]
The researchers are considering all of these possibilities. After considering various possible theories as well as some non-experimental data (including administrative data and data from a special-purpose survey of Migration Organizers who worked on the program in 2017), they feel that the ‘mistargeting’ theory is the most likely explanation and the explanation most consistent with the analysis.[7]
In scenario (1), No Lean Season may be able to identify and fix the problem. In scenario (2), GiveWell will need to update our estimate of the impact of the program to take into account the fact that periodic program failures due to external factors are more likely than we previously thought. In scenario (3), the program is unlikely to be effective in the future.
How did we interpret the RCT results?
We don’t know the extent to which each of the above explanations contributed to the study not finding an effect on migration.
We used the results of the 2017 RCT to update our cost-effectiveness estimate for the program. Cost-effectiveness estimates form arguably the most important single input into our decisions about whether or not to recommend charities (more on how GiveWell uses cost-effectiveness analyses here). When we calculate a program’s cost-effectiveness, we take many different factors into account, such as the administrative and program costs and the expected impact. We also make a number of educated guesses, such as the likelihood that a program’s impact in a new country will be similar to that in a country where it has previously worked. Below, we describe the mechanism by which the 2017 RCT result was incorporated into our model and how it changed our conclusion.
Prior to this year, we formed our view of No Lean Season based on the three small-scale RCTs mentioned above (conducted in 2008, 2013, and 2014). Each of these RCTs looked at a slightly different version of the program. We believed that the ‘high-intensity’ arm of the 2014 RCT was the version most likely to resemble the program at scale. We thus used the migration rate measured in this arm of the RCT as our starting point for calculating the program’s impact.
The high-intensity arm of the 2014 RCT also had the highest measured migration rate of the three RCTs we assessed, and so we wanted to give some consideration to the less-positive results found in the other two assessments. We applied a small, downward adjustment to the rate of induced migration observed in the 2014 high-intensity arm in our cost-effectiveness model; this was an educated guess, based on the information we had. Our best guess was that the program would lead, in expectation, to 80% of the induced migration seen in the 2014 high-intensity arm.[8]
Now, the preliminary 2017 RCT results show no significant impact on migration rates or incomes. Because this trial was large and very recent, we updated our expectations of the impact of the program substantially, and in a negative direction. Our best guess now is that the program will lead, in expectation, to 40% of the induced migration seen in the 2014 high-intensity arm. Holding other inputs constant, this adjustment reduces our estimate of No Lean Season’s cost-effectiveness by a factor of two.
This reduced cost-effectiveness, along with our updated qualitative picture of No Lean Season’s evidence of effectiveness, led to the decision to remove No Lean Season from our top charities list.
What does the future of No Lean Season look like?
Although they are not raising more funding at this time, No Lean Season has over two years’ worth of remaining funding. We understand that the organization has made changes to the program design in 2018 based on emerging interpretations of the 2017 results, and has collected additional data to evaluate some of the hypotheses which may explain those results (including, for example, a survey of Migration Organizers who worked on the 2017 program). They plan to subject the 2018 implementation round to an additional ‘RCT-at-scale,’ with a particular focus on reassessing the program’s effects on migration, income and consumption, as well as potential effects at migration destinations. They will continue to explore what may have caused the issue in the 2017 program at scale, and to see whether they can find a solution. If they do that, we’ll want to reassess the evidence and the costs to determine whether No Lean Season meets our bar for top charity status. Evidence Action believes we should have the necessary information to reassess starting in mid-2019, based on the results of the RCT conducted during the 2018 lean season and other analyses they perform.
Conclusion
This is the second time since 2011 that we have removed a top charity from our list (prior to 2011, our top charities list was fairly different from today; we made a big-picture shift in our priorities that year that led us to our more recent lists). The previous removal occurred in 2013, when we took the Against Malaria Foundation (AMF) off of our list because we didn’t believe it could absorb additional funding effectively in the near term. AMF was reinstated as a top charity in 2014.
The decision to remove a top charity is never easy. But continuously evaluating GiveWell’s recommended charities is an important part of our work, and we take it seriously. It’s easy to talk about a commitment to evidence when the results are positive. It’s hard to maintain that commitment when the results are not. We’re excited to work with a group like Evidence Action that is committed to rigorous program evaluation and open discussion of the results of those evaluations. Its openness about these results has increased our confidence in Evidence Action as an organization. We look forward to seeing the results from the 2018 RCT in 2019.
Notes
[1] “At this early stage in analysis, we find no evidence that the program had an impact (positive or negative) on migration, caloric intake, food expenditure, or income.” Evidence Action, unpublished summary document, Page 1.
[2] The 2017 RCT studied a period from the fall of 2017 through early 2018.
[3] “This study has two main goals:
- “A replication of previous findings showing positive impact of incentivized migration on seasonal migration, caloric intake, food and non-food expenditure, income, and food security. Our aim is to estimate impact of a scaled version of the No Lean Season program: intensifying program implementation within branches and expanding the provision of loans to all eligible households.”
Unpublished summary document, Page 1.
[4] “The second set of explanations focus on unintentional implementation changes caused by the change ineligibility, the vastly expanded scope of the program, or other factors. In the most recent round, it is possible that Migration Organizers (MOs) focused their efforts on those households who were most likely to migrate even without a loan to the exclusion of the target population households who need a loan to afford migration. Such behavior may have even been encouraged by the use of targets set by the NGO to manage implementation at such a large scale. We have implemented a qualitative survey to understand the incentives and actions of MOs last year, and are revising our instructions to avoid any possibility of this issue this year.” Evidence Action, unpublished summary document (with minor revision from Evidence Action), Page 11.
[5] “Most notably, the program was affected by severe flooding in many regions, and implementation was subsequently delayed as well. We are still evaluating whether these regions are the ones with the most diminished effects, although we lack the data in control areas to conduct an experimental comparison.” Evidence Action, unpublished summary document, Page 11-12.
[6] “It is possible that what we observe this year may be the true effect of the No Lean Season program when implemented at scale. This may be because conditions in rural Bangladesh have changed since the initial years of success, spillovers at scale cancel out any gains observed in small-scale pilots, or pilot villages were selected because they were most likely to be receptive to the program.” Evidence Action, unpublished summary document, Page 11.
[7] Evidence Action, “Interpretation of 2017 Results” deck and narrative (unpublished)
[8] “This adjustment is used to account for external validity concerns not accounted for elsewhere in the CEA.
“The default adjustment value of 80% is our best guess about the appropriate value, but it is not based on a formal calculation.
“The program at scale takes place in the same region with the same implementers (RDRS and Evidence Action) as the source of our key evidence for the intervention (the 2014 RCT). The program at scale differs in some aspects of implementation, particularly the inclusiveness of the eligibility criteria and the proportion of eligible households offered an incentive. In the 2014 RCT, the subsidy was a cash transfer rather than an interest-free loan, however the 2008 RCT found a similar effect regardless of whether the subsidy was a cash transfer or an interest-free loan.
“There is some evidence (from a 2013 RCT) suggesting that the program may be ineffective when the perceived risk of migrating increases for reasons such as labor strikes and violence. The researchers estimated that these are 1-in-10 year events.
“Additional discussion related to this parameter can be found at https://www.givewell.org/charities/no-lean-season#programdifferentfromRCTs.” 2018 GiveWell Cost-Effectiveness Model — Version 10, “Migration subsidies” tab, note on cell A19.