Aggregator

In August 2017, GiveWell recommended a grant of $1.3 million to the Centre for Pesticide Suicide Prevention (CPSP). This grant was made as part of GiveWell’s Incubation Grants program to seed the development of potential future GiveWell top charities and to grow the pipeline of organizations we can consider for a recommendation. CPSP implements a different type of program from work GiveWell has funded in the past. Namely, CPSP identifies the pesticides which are most commonly used in suicides and advocates for governments to ban the most lethal pesticides.

Because CPSP's goal is to encourage governments to enact bans, its work falls into the broader category of policy advocacy, an area we are newly focused on. We plan to investigate or are in the process of investigating several other policy causes, including tobacco control, lead paint regulation, and measures to improve road traffic safety.

Summary

This post will discuss:

• GiveWell's interest in researching policy advocacy interventions as possible priority programs. (More)
• Why CPSP is promising as a policy advocacy organization and Incubation Grant recipient. (More)
• Our plans for following CPSP's work going forward. (More)

Read More

The post Considering policy advocacy organizations: Why GiveWell made a grant to the Centre for Pesticide Suicide Prevention appeared first on The GiveWell Blog.

In August 2017, GiveWell recommended a grant of $1.3 million to the Centre for Pesticide Suicide Prevention (CPSP). This grant was made as part of GiveWell’s Incubation Grants program to seed the development of potential future GiveWell top charities and to grow the pipeline of organizations we can consider for a recommendation. CPSP implements a different type of program from work GiveWell has funded in the past. Namely, CPSP identifies the pesticides which are most commonly used in suicides and advocates for governments to ban the most lethal pesticides.

Because CPSP’s goal is to encourage governments to enact bans, its work falls into the broader category of policy advocacy, an area we are newly focused on. We plan to investigate or are in the process of investigating several other policy causes, including tobacco control, lead paint regulation, and measures to improve road traffic safety.

Summary

This post will discuss:

• GiveWell’s interest in researching policy advocacy interventions as possible priority programs. (More)
• Why CPSP is promising as a policy advocacy organization and Incubation Grant recipient. (More)
• Our plans for following CPSP’s work going forward. (More)

Policy advocacy work

One of the key criteria we use to evaluate potential top charities is their cost-effectiveness—how much good each dollar donated to that charity can accomplish. In recent years, we’ve identified several charities that we estimate to be around 4 to 10 times as cost-effective as GiveDirectly, which we use as a benchmark for cost-effectiveness. Our top charities are extremely cost-effective, but we wonder whether we might be able to find opportunities that are significantly more cost-effective than the charities we currently recommend.

Our current top charities largely focus on direct implementation of health and poverty alleviation interventions. One of our best guesses for where we might find significantly more cost-effective charities is in the area of policy advocacy, or programs that aim to influence government policy. Our intuition is that spending a relatively small amount of money on advocacy could lead to policy changes resulting in long-run benefits for many people, and thus could be among the most cost-effective ways to help people. As a result, researching policy advocacy interventions is one of our biggest priorities for the year ahead.

Policy advocacy work may have the following advantages:

• Leverage: A relatively small amount of spending on advocacy may influence larger amounts of government funding;
• Sustainability: A policy may be in place for years after its adoption; and
• Feasibility: Some effective interventions can only be effectively implemented by governments, such as increasing taxes on tobacco to reduce consumption.

Policy advocacy also poses serious challenges for GiveWell when we consider it as a potential priority area:

• Evidence of effectiveness will likely be lower quality than what we’ve seen from our top charities, e.g. it may involve analyzing trends over time (where confounding factors may complicate analysis) rather than randomized controlled trials or quasi-experimental evidence;
• Causal attribution will be challenging in that multiple players are likely to be involved in any policy change and policymakers are likely to be influenced by a variety of factors;
• There may be a substantial chance of failure to pass the desired legislation; and
• Regulation may have undesirable secondary effects.

Overall, evaluating policy advocacy requires a different approach to assessing evidence and probability of success than our top charities work has in the past.

Incubation Grant to the Centre for Pesticide Suicide Prevention

CPSP began work in 2016 and aims to reduce deaths due to deliberate ingestion of lethal pesticides. With this Incubation Grant, which is intended to cover two years of expenses, CPSP expects to collect data on which pesticides are most often used in suicide attempts and which are most lethal, and then to use this data to advocate to the governments of India and Nepal to implement bans of certain lethal pesticides.

Research suggests that worldwide, approximately 14% to 20% of suicides involved the deliberate ingestion of pesticides. This method of suicide may be particularly common in agricultural populations. The case we see for this grant relies largely on data from Sri Lanka, where bans on the pesticides that were most lethal and most commonly used in suicide coincided with a substantial decrease in the overall suicide rate; we find the case that the decline in suicides was primarily caused by the pesticide bans reasonably compelling. CPSP’s director, Michael Eddleston, was involved in advocating for some of those bans. Read more here.

GiveWell learned of CPSP’s work through James Snowden, who joined GiveWell as a Research Consultant in early 2017. We decided to recommend support to CPSP based on the evidence that pesticide regulation may reduce overall suicide rates, our impression that an advocacy organization could effect changes in regulations, our view that Michael Eddleston and Leah Utyasheva (the co-founders) are well-positioned to do this type of work, and our expectation that we would be able to evaluate CPSP’s impact on pesticide regulation in Nepal and India over the next few years. We thus think CPSP is a plausible future GiveWell top charity and a good fit for an Incubation Grant.

While deciding whether to make this grant, GiveWell staff discussed how to think about the impact of preventing a suicide. Thinking about this question depends on limited empirical information, and staff did not come to an internal consensus. Our best guess at this point is that CPSP generally prevents suicide by people who are making impulsive decisions.

We see several risks to the success of this grant:

• Banning lethal pesticides may be ineffective as a means of preventing suicide, in India and Nepal or more broadly. The case for this area of policy advocacy relies largely on the observational studies from Sri Lanka mentioned above, supported by Sri Lankan medical records suggesting the decline is partially explained by a shift to less lethal pesticides in suicide attempts.
• CPSP may not be able to translate its research into policy change. This risk of failure to achieve legislative change characterizes policy advocacy work in general, to some extent, and requires us to make a type of prediction that is not needed when evaluating a charity directly implementing a program.
• Banning pesticides could lead to offsetting effects in agricultural production. The limited evidence we have seen on this question suggests that past pesticide bans have not led to notable decreases in agricultural production, but we still believe this is a risk.
• CPSP is a new organization, so it does not have a track record of successfully conducting this type of research and achieving policy change.

To quantify the risks above, GiveWell Executive Director Elie Hassenfeld and James Snowden each recorded predictions about the outcomes of this grant at the time the grant was made. Briefly (more predictions here), Elie and James predict with 33% and 55% probability, respectively, that Nepal will pass legislation banning at least one of the three pesticides most commonly used in suicide by July 1, 2020, and with 15% and 35% probability, respectively, that at least one state in India will do so.

Going forward

We plan to continue having regular conversations with CPSP, and a more substantial check-in one year after the grant was made. At that point, we intend to assess whether CPSP has been meeting the milestones it expected to meet and decide whether to provide a third year of funding. If this grant is successful, we hope we may be able to evaluate CPSP as a potential top charity.

The post Considering policy advocacy organizations: Why GiveWell made a grant to the Centre for Pesticide Suicide Prevention appeared first on The GiveWell Blog.

In August 2017, GiveWell recommended a grant of $1.3 million to the Centre for Pesticide Suicide Prevention (CPSP). This grant was made as part of GiveWell’s Incubation Grants program to seed the development of potential future GiveWell top charities and to grow the pipeline of organizations we can consider for a recommendation. CPSP implements a different type of program from work GiveWell has funded in the past. Namely, CPSP identifies the pesticides which are most commonly used in suicides and advocates for governments to ban the most lethal pesticides.

Because CPSP’s goal is to encourage governments to enact bans, its work falls into the broader category of policy advocacy, an area we are newly focused on. We plan to investigate or are in the process of investigating several other policy causes, including tobacco control, lead paint regulation, and measures to improve road traffic safety.

Summary

This post will discuss:

• GiveWell’s interest in researching policy advocacy interventions as possible priority programs. (More)
• Why CPSP is promising as a policy advocacy organization and Incubation Grant recipient. (More)
• Our plans for following CPSP’s work going forward. (More)

Policy advocacy work

One of the key criteria we use to evaluate potential top charities is their cost-effectiveness—how much good each dollar donated to that charity can accomplish. In recent years, we’ve identified several charities that we estimate to be around 4 to 10 times as cost-effective as GiveDirectly, which we use as a benchmark for cost-effectiveness. Our top charities are extremely cost-effective, but we wonder whether we might be able to find opportunities that are significantly more cost-effective than the charities we currently recommend.

Our current top charities largely focus on direct implementation of health and poverty alleviation interventions. One of our best guesses for where we might find significantly more cost-effective charities is in the area of policy advocacy, or programs that aim to influence government policy. Our intuition is that spending a relatively small amount of money on advocacy could lead to policy changes resulting in long-run benefits for many people, and thus could be among the most cost-effective ways to help people. As a result, researching policy advocacy interventions is one of our biggest priorities for the year ahead.

Policy advocacy work may have the following advantages:

• Leverage: A relatively small amount of spending on advocacy may influence larger amounts of government funding;
• Sustainability: A policy may be in place for years after its adoption; and
• Feasibility: Some effective interventions can only be effectively implemented by governments, such as increasing taxes on tobacco to reduce consumption.

Policy advocacy also poses serious challenges for GiveWell when we consider it as a potential priority area:

• Evidence of effectiveness will likely be lower quality than what we’ve seen from our top charities, e.g. it may involve analyzing trends over time (where confounding factors may complicate analysis) rather than randomized controlled trials or quasi-experimental evidence;
• Causal attribution will be challenging in that multiple players are likely to be involved in any policy change and policymakers are likely to be influenced by a variety of factors;
• There may be a substantial chance of failure to pass the desired legislation; and
• Regulation may have undesirable secondary effects.

Overall, evaluating policy advocacy requires a different approach to assessing evidence and probability of success than our top charities work has in the past.

Incubation Grant to the Centre for Pesticide Suicide Prevention

CPSP began work in 2016 and aims to reduce deaths due to deliberate ingestion of lethal pesticides. With this Incubation Grant, which is intended to cover two years of expenses, CPSP expects to collect data on which pesticides are most often used in suicide attempts and which are most lethal, and then to use this data to advocate to the governments of India and Nepal to implement bans of certain lethal pesticides.

Research suggests that worldwide, approximately 14% to 20% of suicides involved the deliberate ingestion of pesticides. This method of suicide may be particularly common in agricultural populations. The case we see for this grant relies largely on data from Sri Lanka, where bans on the pesticides that were most lethal and most commonly used in suicide coincided with a substantial decrease in the overall suicide rate; we find the case that the decline in suicides was primarily caused by the pesticide bans reasonably compelling. CPSP’s director, Michael Eddleston, was involved in advocating for some of those bans. Read more here.

GiveWell learned of CPSP’s work through James Snowden, who joined GiveWell as a Research Consultant in early 2017. We decided to recommend support to CPSP based on the evidence that pesticide regulation may reduce overall suicide rates, our impression that an advocacy organization could effect changes in regulations, our view that Michael Eddleston and Leah Utyasheva (the co-founders) are well-positioned to do this type of work, and our expectation that we would be able to evaluate CPSP’s impact on pesticide regulation in Nepal and India over the next few years. We thus think CPSP is a plausible future GiveWell top charity and a good fit for an Incubation Grant.

While deciding whether to make this grant, GiveWell staff discussed how to think about the impact of preventing a suicide. Thinking about this question depends on limited empirical information, and staff did not come to an internal consensus. Our best guess at this point is that CPSP generally prevents suicide by people who are making impulsive decisions.

We see several risks to the success of this grant:

• Banning lethal pesticides may be ineffective as a means of preventing suicide, in India and Nepal or more broadly. The case for this area of policy advocacy relies largely on the observational studies from Sri Lanka mentioned above, supported by Sri Lankan medical records suggesting the decline is partially explained by a shift to less lethal pesticides in suicide attempts.
• CPSP may not be able to translate its research into policy change. This risk of failure to achieve legislative change characterizes policy advocacy work in general, to some extent, and requires us to make a type of prediction that is not needed when evaluating a charity directly implementing a program.
• Banning pesticides could lead to offsetting effects in agricultural production. The limited evidence we have seen on this question suggests that past pesticide bans have not led to notable decreases in agricultural production, but we still believe this is a risk.
• CPSP is a new organization, so it does not have a track record of successfully conducting this type of research and achieving policy change.

To quantify the risks above, GiveWell Executive Director Elie Hassenfeld and James Snowden each recorded predictions about the outcomes of this grant at the time the grant was made. Briefly (more predictions here), Elie and James predict with 33% and 55% probability, respectively, that Nepal will pass legislation banning at least one of the three pesticides most commonly used in suicide by July 1, 2020, and with 15% and 35% probability, respectively, that at least one state in India will do so.

Going forward

We plan to continue having regular conversations with CPSP, and a more substantial check-in one year after the grant was made. At that point, we intend to assess whether CPSP has been meeting the milestones it expected to meet and decide whether to provide a third year of funding. If this grant is successful, we hope we may be able to evaluate CPSP as a potential top charity.

The post Considering policy advocacy organizations: Why GiveWell made a grant to the Centre for Pesticide Suicide Prevention appeared first on The GiveWell Blog.

Our goal with hosting quarterly open threads is to give blog readers an opportunity to publicly raise comments or questions about GiveWell or related topics (in the comments section below). As always, you’re also welcome to email us at info@givewell.org or to request a call with GiveWell staff if you have feedback or questions you’d prefer to discuss privately. We’ll try to respond promptly to questions or comments.

You can view our December 2017 open thread here.

The post March 2018 open thread appeared first on The GiveWell Blog.

Our goal with hosting quarterly open threads is to give blog readers an opportunity to publicly raise comments or questions about GiveWell or related topics (in the comments section below). As always, you’re also welcome to email us at info@givewell.org or to request a call with GiveWell staff if you have feedback or questions you’d prefer to discuss privately. We’ll try to respond promptly to questions or comments.

You can view our December 2017 open thread here.

The post March 2018 open thread appeared first on The GiveWell Blog.

Many charities aim to influence how others (other donors, governments, or the private sector) allocate their funds. We call this influence on others “leverage.” Expenditure on a program can also crowd out funding that would otherwise have come from other sources. We call this “funging” (from “fungibility”).

In GiveWell’s early years, we didn’t account for leverage in our cost-effectiveness analysis; we counted all costs of an intervention equally, no matter who paid for them.1For example, see row 3 of our 2013 cost-effectiveness analysis for Against Malaria Foundation. jQuery("#footnote_plugin_tooltip_1").tooltip({ tip: "#footnote_plugin_tooltip_text_1", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); For example, for the Schistosomiasis Control Initiative (SCI), a charity that treats intestinal parasites (deworming), we counted both drug and delivery costs, even when the drugs were donated. We did this because we felt it was the simplest approach, least prone to significant error or manipulation.

Over the last few years, our approach has evolved, and we made some adjustments for leverage and funging to our cost-effectiveness analyses where we felt they were clearly warranted.

In our top charities update at the end of 2017, we made a major change to how we dealt with the question of leverage by incorporating explicit, formal leverage estimates for every charity we recommend.

This change made our cost-effectiveness estimates of deworming charities (which typically leverage substantial government funding) look more cost-effective than our previous method. For example, our new method makes SCI look 1.2x more cost-effective than in the previous cost-effectiveness update. More details are in the table at the end of this post.

We also think the change makes our reasoning more transparent and more consistent across organizations.

In this post, we:

• Describe how our treatment of leverage and funging has evolved.
• Highlight two major limitations of our current approach.
• Present how much difference leverage and funging make to our cost-effectiveness estimates.

Details follow.

How our thinking has evolved

We last wrote about our approach to leverage and funging in a 2011 blog post. In short, we didn’t explicitly account for leverage in our cost-effectiveness analysis, counting costs to all entities equally. We concluded:

When we do cost-effectiveness estimates (e.g., “cost per life saved”) we consider all expenses from all sources, not just funding provided by GiveWell donors. For SCI, we count both drug and delivery costs, even when drugs are donated. (Generally, we try to count all donated goods and services at market value, i.e., the price the donor could have sold them for instead of donating them.) For [the Against Malaria Foundation (AMF)], we count net costs and distribution costs, even though AMF pays only for the former. In the case of VillageReach, we even count government costs of delivering vaccines, even though VillageReach works exclusively to improve the efficiency of the delivery system.

We consider this approach the simplest approach to dealing with the issues discussed here, and given our limited understanding of how “leverage” works, we believe that this approach minimizes the error in our estimates that might come from misreading the “leverage” situation. As our understanding of “leverage” improves, we may approach our cost-effectiveness estimates differently.

Since 2011, our thinking changed. Over time, we started applying some adjustments to our cost-effectiveness model to account for leverage and funging when it seemed important to our bottom line and fairly clear that some adjustment was warranted:

• We applied discounts to costs incurred by certain entities. For example, we applied a 50% discount to the value of teacher time spent distributing deworming tablets, and excluded the costs to pharmaceutical companies donating these tablets.2See our May 2017 cost-effectiveness analysis. jQuery("#footnote_plugin_tooltip_2").tooltip({ tip: "#footnote_plugin_tooltip_text_2", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); Our rationale was that without our top charities, these resources would likely otherwise been used less productively.
• We applied ‘alternative funders adjustments’ to account for the possibility that we were crowding out other funders. For example, some of the distributions that AMF considered funding, but didn’t ultimately fund, were picked up by other funders (more).

This helped us explicitly think through considerations relevant to our top charities. But by the end of 2016, our model had a handful of ad hoc adjustments that were difficult to identify, understand, and vet. For example, the discounts we applied to costs incurred by certain entities were ‘baked in’ to our estimates of cost per treatment, rather than explicit on the main spreadsheet of our cost-effectiveness analysis.

Changes to how we incorporate leverage and funging into our cost-effectiveness analysis

We revisited the way we thought about leverage and funging in preparation for our 2017 top charities decision. We wanted to make sure our adjustments were transparent and consistent across all charities.

We now explicitly make quantitative judgments about (i) the probability that our charities are causing governments and multilateral aid agencies to spend more or less on a program than they otherwise would have and (ii) the value of what those funds would otherwise have been spent on.3Our current best guess of a reasonable benchmark for the counterfactual value of government funds is ~75% as cost-effective as GiveDirectly (discussed later in the post). We view this is a very rough guess. jQuery("#footnote_plugin_tooltip_3").tooltip({ tip: "#footnote_plugin_tooltip_text_3", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

Here’s an exercise that some GiveWell staff have found helpful for getting a more intuitive feel for different ways of treating leverage.

Suppose a charity pays $5,000 to purchase magic pills. This would cause (with 100% certainty) the government to spend another $5,000 distributing those pills. The pill distribution saves 1,000 lives in total. If the government didn’t fund the pill distribution, it would have spent $5,000 on something that would have saved 250 lives.

How should a philanthropist think about the cost-effectiveness of this charity?

1. One option is to include all costs to all actors on the cost side of the cost-effectiveness ratio. Total costs are $10,000 to save 1,000 lives and cost-effectiveness is $10 / life saved. This was GiveWell’s approach in 2011.
2. Another option is to discount government costs by 50%, because the government would otherwise have spent the funds on something 50% as effective. So total costs are $5,000 + (50% x $5,000) = $7,500. 1,000 lives are saved and cost-effectiveness is $7.50 / life saved. This was GiveWell’s approach from 2014 through 2016.
3. A third option is to include only the costs to the charity on the ‘cost’ side. The charity causes the magic pill distribution to happen, saving 1,000 lives. But it also causes the government to spend $5,000, which otherwise would have been used to save 250 lives. So the total costs are $5,000, and 1,000 – 250 = 750 lives are saved. Cost-effectiveness is $6.66 / life saved. This is GiveWell’s approach now.4In order to isolate the effect that leverage/funging has, we first calculate the impact of the program using the first method (including all costs equally), then apply a “leverage/funging” adjustment to transform the answer to the third method. jQuery("#footnote_plugin_tooltip_4").tooltip({ tip: "#footnote_plugin_tooltip_text_4", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

We believe the third way of treating leverage best reflects the true counterfactual impact of a charity’s activities. It also makes charities that are leveraging other funders look substantially more cost-effective than we previously thought.

Limitations of our approach

There are two important limitations to the way we account for leverage and funging.

First, these estimates rely on more guesswork than most of our cost-effectiveness analysis, reflecting a fundamental tradeoff we face in deciding which considerations to explicitly quantify. Quantification forces us to think through not just whether a particular consideration matters, but how much it matters relative to other factors, and to be explicit about that. On the other hand, incorporating very uncertain factors into our analysis can reduce its reliability, give a false impression of certainty, and make it difficult for others to engage with our work. In this case, we thought the benefits of explicit quantification outweighed the costs.

Two examples of assumptions going into our leverage and funging adjustments that we’re highly uncertain about:

1. Our best guess is that the average counterfactual use of domestic government spending that could be leveraged by our top charities is ~75% as cost-effective as GiveDirectly. We think using this figure is a useful heuristic, which roughly accords with our intuitions (and ensures we’re being consistent between charities), but we don’t feel confident that we have a good sense of what governments would counterfactually spend their funds on, or how valuable those activities might be.
2. We estimate there is a ~70% chance that, without Malaria Consortium funding, the marginal seasonal malaria chemoprevention (SMC) program would go unfunded, but only a ~40% chance that, without Against Malaria Foundation funding, the marginal bednet distribution would go unfunded. Estimating these probabilities is challenging, but taking our best guess forces us to evaluate how much weight to place on the qualitative consideration that there are more alternative funders for bednet distribution than SMC.

Second, we don’t explicitly model the long-term financial sustainability of a program. One worldview we find plausible for the role of effective philanthropy is in demonstrating the effectiveness of novel projects that, in the long run, are taken up by governments. This is not captured within our current model, which only looks at the effects of leverage and funging in the short term. Due to the difficulty of explicitly modelling this consideration, we take it into account qualitatively.5For example, we allocated more discretionary funding than we would have on the basis of cost-effectiveness alone to No Lean Season in 2017 due to our view that it was demonstrating the effectiveness of a novel program, which may have long-run funding implications. jQuery("#footnote_plugin_tooltip_5").tooltip({ tip: "#footnote_plugin_tooltip_text_5", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

 

How much of a difference do leverage and funging make?

In the table below, we present how our new method of accounting for leverage and funging compares to (i) counting all costs equally and (ii) our previous method of accounting for leverage and funging.

Adjustments range between a modest penalty for AMF (because we expect AMF crowds out some funds from other sources) to a large boost to SCI (because the cost to pharmaceutical companies of manufacturing donated drugs comprises a substantial proportion of cost per treatment in SCI distributions, and we expect that without SCI, these resources would have been put to less valuable uses).

Note: 1.2x implies the adjustment makes the charity look 20% more cost-effective; 0.8x implies the adjustment makes the charity look 20% less cost-effective. All charities listed are GiveWell top charities as of November 2017.

Charity Versus counting all costs equally6Calculations here. “N/A” refers to charities for which we had not completed a cost-effectiveness analysis before October 2017. jQuery("#footnote_plugin_tooltip_6").tooltip({ tip: "#footnote_plugin_tooltip_text_6", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); Versus our 2014-16 methodology Commentary Against Malaria Foundation 0.8x 1.1x Government costs represent a small proportion of funding for AMF programs. Our analysis of distributions that AMF considered, but did not fund, suggests that some of these distributions are covered by alternative funders, who would otherwise have supported less valuable programs. Schistosomiasis Control Initiative 2x 1.2x We estimate ~60% of the costs of SCI-supported deworming programs are incurred by either governments or pharmaceutical companies. We expect that without SCI, most of these resources would have been used on less valuable programs. Evidence Action’s Deworm the World Initiative 1.4x 1.1x We estimate ~40% of the costs of Deworm the World-supported deworming programs are incurred by either governments or pharmaceutical companies. We expect that without Deworm the World, most of these resources would have been used on less valuable programs. Sightsavers’ deworming program 1.6x 1.3x We estimate ~50% of the costs of deworming in Sightsavers supported programs are from governments or donated drugs from pharmaceutical companies. We expect that without Sightsavers, most of these resources would have been used on less valuable programs. END Fund’s deworming program 1.3x N/A We estimate ~40% of the costs of END Fund-supported deworming programs are incurred by either governments or pharmaceutical companies. We expect that without the END Fund, most of these resources would have been used on less valuable programs. Helen Keller International (HKI)’s vitamin A supplementation (VAS) program 1.1x N/A We estimate ~25% of the costs of HKI-supported VAS programs are covered by governments. We expect that without HKI, most of these resources would have been used on less valuable programs. GiveDirectly 1x 1x Due to the scalability of GiveDirectly’s program, we believe it is unlikely that GiveDirectly crowds out funding from other sources. GiveDirectly does not leverage funds from other sources. Malaria Consortium’s seasonal malaria chemoprevention program .98x 1.04x Government costs represent a small proportion of funding for Malaria Consortium programs. We believe it is possible but unlikely that Malaria Consortium crowds out additional government funding. Evidence Action’s No Lean Season 1x N/A No Lean Season is a novel program, and we think it’s unlikely to be crowding out funding from other sources. No Lean Season does not leverage substantial funding from other sources.

Notes   [ + ]

1. ↑ For example, see row 3 of our 2013 cost-effectiveness analysis for Against Malaria Foundation. 2. ↑ See our May 2017 cost-effectiveness analysis. 3. ↑ Our current best guess of a reasonable benchmark for the counterfactual value of government funds is ~75% as cost-effective as GiveDirectly (discussed later in the post). We view this is a very rough guess. 4. ↑ In order to isolate the effect that leverage/funging has, we first calculate the impact of the program using the first method (including all costs equally), then apply a “leverage/funging” adjustment to transform the answer to the third method. 5. ↑ For example, we allocated more discretionary funding than we would have on the basis of cost-effectiveness alone to No Lean Season in 2017 due to our view that it was demonstrating the effectiveness of a novel program, which may have long-run funding implications. 6. ↑ Calculations here. “N/A” refers to charities for which we had not completed a cost-effectiveness analysis before October 2017. function footnote_expand_reference_container() { jQuery("#footnote_references_container").show(); jQuery("#footnote_reference_container_collapse_button").text("-"); } function footnote_collapse_reference_container() { jQuery("#footnote_references_container").hide(); jQuery("#footnote_reference_container_collapse_button").text("+"); } function footnote_expand_collapse_reference_container() { if (jQuery("#footnote_references_container").is(":hidden")) { footnote_expand_reference_container(); } else { footnote_collapse_reference_container(); } } function footnote_moveToAnchor(p_str_TargetID) { footnote_expand_reference_container(); var l_obj_Target = jQuery("#" + p_str_TargetID); if(l_obj_Target.length) { jQuery('html, body').animate({ scrollTop: l_obj_Target.offset().top - window.innerHeight/2 }, 1000); } }

The post Revisiting leverage appeared first on The GiveWell Blog.

Many charities aim to influence how others (other donors, governments, or the private sector) allocate their funds. We call this influence on others “leverage.” Expenditure on a program can also crowd out funding that would otherwise have come from other sources. We call this “funging” (from “fungibility”).

In GiveWell’s early years, we didn’t account for leverage in our cost-effectiveness analysis; we counted all costs of an intervention equally, no matter who paid for them.1For example, see row 3 of our 2013 cost-effectiveness analysis for Against Malaria Foundation. jQuery("#footnote_plugin_tooltip_1").tooltip({ tip: "#footnote_plugin_tooltip_text_1", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); For example, for the Schistosomiasis Control Initiative (SCI), a charity that treats intestinal parasites (deworming), we counted both drug and delivery costs, even when the drugs were donated. We did this because we felt it was the simplest approach, least prone to significant error or manipulation.

Over the last few years, our approach has evolved, and we made some adjustments for leverage and funging to our cost-effectiveness analyses where we felt they were clearly warranted.

In our top charities update at the end of 2017, we made a major change to how we dealt with the question of leverage by incorporating explicit, formal leverage estimates for every charity we recommend.

This change made our cost-effectiveness estimates of deworming charities (which typically leverage substantial government funding) look more cost-effective than our previous method. For example, our new method makes SCI look 1.2x more cost-effective than in the previous cost-effectiveness update. More details are in the table at the end of this post.

We also think the change makes our reasoning more transparent and more consistent across organizations.

In this post, we:

• Describe how our treatment of leverage and funging has evolved.
• Highlight two major limitations of our current approach.
• Present how much difference leverage and funging make to our cost-effectiveness estimates.

Details follow.

How our thinking has evolved

We last wrote about our approach to leverage and funging in a 2011 blog post. In short, we didn’t explicitly account for leverage in our cost-effectiveness analysis, counting costs to all entities equally. We concluded:

When we do cost-effectiveness estimates (e.g., “cost per life saved”) we consider all expenses from all sources, not just funding provided by GiveWell donors. For SCI, we count both drug and delivery costs, even when drugs are donated. (Generally, we try to count all donated goods and services at market value, i.e., the price the donor could have sold them for instead of donating them.) For [the Against Malaria Foundation (AMF)], we count net costs and distribution costs, even though AMF pays only for the former. In the case of VillageReach, we even count government costs of delivering vaccines, even though VillageReach works exclusively to improve the efficiency of the delivery system.

We consider this approach the simplest approach to dealing with the issues discussed here, and given our limited understanding of how “leverage” works, we believe that this approach minimizes the error in our estimates that might come from misreading the “leverage” situation. As our understanding of “leverage” improves, we may approach our cost-effectiveness estimates differently.

Since 2011, our thinking changed. Over time, we started applying some adjustments to our cost-effectiveness model to account for leverage and funging when it seemed important to our bottom line and fairly clear that some adjustment was warranted:

• We applied discounts to costs incurred by certain entities. For example, we applied a 50% discount to the value of teacher time spent distributing deworming tablets, and excluded the costs to pharmaceutical companies donating these tablets.2See our May 2017 cost-effectiveness analysis. jQuery("#footnote_plugin_tooltip_2").tooltip({ tip: "#footnote_plugin_tooltip_text_2", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); Our rationale was that without our top charities, these resources would likely otherwise been used less productively.
• We applied ‘alternative funders adjustments’ to account for the possibility that we were crowding out other funders. For example, some of the distributions that AMF considered funding, but didn’t ultimately fund, were picked up by other funders (more).

This helped us explicitly think through considerations relevant to our top charities. But by the end of 2016, our model had a handful of ad hoc adjustments that were difficult to identify, understand, and vet. For example, the discounts we applied to costs incurred by certain entities were ‘baked in’ to our estimates of cost per treatment, rather than explicit on the main spreadsheet of our cost-effectiveness analysis.

Changes to how we incorporate leverage and funging into our cost-effectiveness analysis

We revisited the way we thought about leverage and funging in preparation for our 2017 top charities decision. We wanted to make sure our adjustments were transparent and consistent across all charities.

We now explicitly make quantitative judgments about (i) the probability that our charities are causing governments and multilateral aid agencies to spend more or less on a program than they otherwise would have and (ii) the value of what those funds would otherwise have been spent on.3Our current best guess of a reasonable benchmark for the counterfactual value of government funds is ~75% as cost-effective as GiveDirectly (discussed later in the post). We view this is a very rough guess. jQuery("#footnote_plugin_tooltip_3").tooltip({ tip: "#footnote_plugin_tooltip_text_3", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

Here’s an exercise that some GiveWell staff have found helpful for getting a more intuitive feel for different ways of treating leverage.

Suppose a charity pays $5,000 to purchase magic pills. This would cause (with 100% certainty) the government to spend another $5,000 distributing those pills. The pill distribution saves 1,000 lives in total. If the government didn’t fund the pill distribution, it would have spent $5,000 on something that would have saved 250 lives.

How should a philanthropist think about the cost-effectiveness of this charity?

1. One option is to include all costs to all actors on the cost side of the cost-effectiveness ratio. Total costs are $10,000 to save 1,000 lives and cost-effectiveness is $10 / life saved. This was GiveWell’s approach in 2011.
2. Another option is to discount government costs by 50%, because the government would otherwise have spent the funds on something 50% as effective. So total costs are $5,000 + (50% x $5,000) = $7,500. 1,000 lives are saved and cost-effectiveness is $7.50 / life saved. This was GiveWell’s approach from 2014 through 2016.
3. A third option is to include only the costs to the charity on the ‘cost’ side. The charity causes the magic pill distribution to happen, saving 1,000 lives. But it also causes the government to spend $5,000, which otherwise would have been used to save 250 lives. So the total costs are $5,000, and 1,000 – 250 = 750 lives are saved. Cost-effectiveness is $6.66 / life saved. This is GiveWell’s approach now.4In order to isolate the effect that leverage/funging has, we first calculate the impact of the program using the first method (including all costs equally), then apply a “leverage/funging” adjustment to transform the answer to the third method. jQuery("#footnote_plugin_tooltip_4").tooltip({ tip: "#footnote_plugin_tooltip_text_4", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

We believe the third way of treating leverage best reflects the true counterfactual impact of a charity’s activities. It also makes charities that are leveraging other funders look substantially more cost-effective than we previously thought.

Limitations of our approach

There are two important limitations to the way we account for leverage and funging.

First, these estimates rely on more guesswork than most of our cost-effectiveness analysis, reflecting a fundamental tradeoff we face in deciding which considerations to explicitly quantify. Quantification forces us to think through not just whether a particular consideration matters, but how much it matters relative to other factors, and to be explicit about that. On the other hand, incorporating very uncertain factors into our analysis can reduce its reliability, give a false impression of certainty, and make it difficult for others to engage with our work. In this case, we thought the benefits of explicit quantification outweighed the costs.

Two examples of assumptions going into our leverage and funging adjustments that we’re highly uncertain about:

1. Our best guess is that the average counterfactual use of domestic government spending that could be leveraged by our top charities is ~75% as cost-effective as GiveDirectly. We think using this figure is a useful heuristic, which roughly accords with our intuitions (and ensures we’re being consistent between charities), but we don’t feel confident that we have a good sense of what governments would counterfactually spend their funds on, or how valuable those activities might be.
2. We estimate there is a ~70% chance that, without Malaria Consortium funding, the marginal seasonal malaria chemoprevention (SMC) program would go unfunded, but only a ~40% chance that, without Against Malaria Foundation funding, the marginal bednet distribution would go unfunded. Estimating these probabilities is challenging, but taking our best guess forces us to evaluate how much weight to place on the qualitative consideration that there are more alternative funders for bednet distribution than SMC.

Second, we don’t explicitly model the long-term financial sustainability of a program. One worldview we find plausible for the role of effective philanthropy is in demonstrating the effectiveness of novel projects that, in the long run, are taken up by governments. This is not captured within our current model, which only looks at the effects of leverage and funging in the short term. Due to the difficulty of explicitly modelling this consideration, we take it into account qualitatively.5For example, we allocated more discretionary funding than we would have on the basis of cost-effectiveness alone to No Lean Season in 2017 due to our view that it was demonstrating the effectiveness of a novel program, which may have long-run funding implications. jQuery("#footnote_plugin_tooltip_5").tooltip({ tip: "#footnote_plugin_tooltip_text_5", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

 

How much of a difference do leverage and funging make?

In the table below, we present how our new method of accounting for leverage and funging compares to (i) counting all costs equally and (ii) our previous method of accounting for leverage and funging.

Adjustments range between a modest penalty for AMF (because we expect AMF crowds out some funds from other sources) to a large boost to SCI (because the cost to pharmaceutical companies of manufacturing donated drugs comprises a substantial proportion of cost per treatment in SCI distributions, and we expect that without SCI, these resources would have been put to less valuable uses).

Note: 1.2x implies the adjustment makes the charity look 20% more cost-effective; 0.8x implies the adjustment makes the charity look 20% less cost-effective. All charities listed are GiveWell top charities as of November 2017.

Charity Versus counting all costs equally6Calculations here. “N/A” refers to charities for which we had not completed a cost-effectiveness analysis before October 2017. jQuery("#footnote_plugin_tooltip_6").tooltip({ tip: "#footnote_plugin_tooltip_text_6", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); Versus our 2014-16 methodology Commentary Against Malaria Foundation 0.8x 1.1x Government costs represent a small proportion of funding for AMF programs. Our analysis of distributions that AMF considered, but did not fund, suggests that some of these distributions are covered by alternative funders, who would otherwise have supported less valuable programs. Schistosomiasis Control Initiative 2x 1.2x We estimate ~60% of the costs of SCI-supported deworming programs are incurred by either governments or pharmaceutical companies. We expect that without SCI, most of these resources would have been used on less valuable programs. Evidence Action’s Deworm the World Initiative 1.4x 1.1x We estimate ~40% of the costs of Deworm the World-supported deworming programs are incurred by either governments or pharmaceutical companies. We expect that without Deworm the World, most of these resources would have been used on less valuable programs. Sightsavers’ deworming program 1.6x 1.3x We estimate ~50% of the costs of deworming in Sightsavers supported programs are from governments or donated drugs from pharmaceutical companies. We expect that without Sightsavers, most of these resources would have been used on less valuable programs. END Fund’s deworming program 1.3x N/A We estimate ~40% of the costs of END Fund-supported deworming programs are incurred by either governments or pharmaceutical companies. We expect that without the END Fund, most of these resources would have been used on less valuable programs. Helen Keller International (HKI)’s vitamin A supplementation (VAS) program 1.1x N/A We estimate ~25% of the costs of HKI-supported VAS programs are covered by governments. We expect that without HKI, most of these resources would have been used on less valuable programs. GiveDirectly 1x 1x Due to the scalability of GiveDirectly’s program, we believe it is unlikely that GiveDirectly crowds out funding from other sources. GiveDirectly does not leverage funds from other sources. Malaria Consortium’s seasonal malaria chemoprevention program .98x 1.04x Government costs represent a small proportion of funding for Malaria Consortium programs. We believe it is possible but unlikely that Malaria Consortium crowds out additional government funding. Evidence Action’s No Lean Season 1x N/A No Lean Season is a novel program, and we think it’s unlikely to be crowding out funding from other sources. No Lean Season does not leverage substantial funding from other sources.

Notes   [ + ]

1. ↑ For example, see row 3 of our 2013 cost-effectiveness analysis for Against Malaria Foundation. 2. ↑ See our May 2017 cost-effectiveness analysis. 3. ↑ Our current best guess of a reasonable benchmark for the counterfactual value of government funds is ~75% as cost-effective as GiveDirectly (discussed later in the post). We view this is a very rough guess. 4. ↑ In order to isolate the effect that leverage/funging has, we first calculate the impact of the program using the first method (including all costs equally), then apply a “leverage/funging” adjustment to transform the answer to the third method. 5. ↑ For example, we allocated more discretionary funding than we would have on the basis of cost-effectiveness alone to No Lean Season in 2017 due to our view that it was demonstrating the effectiveness of a novel program, which may have long-run funding implications. 6. ↑ Calculations here. “N/A” refers to charities for which we had not completed a cost-effectiveness analysis before October 2017. function footnote_expand_reference_container() { jQuery("#footnote_references_container").show(); jQuery("#footnote_reference_container_collapse_button").text("-"); } function footnote_collapse_reference_container() { jQuery("#footnote_references_container").hide(); jQuery("#footnote_reference_container_collapse_button").text("+"); } function footnote_expand_collapse_reference_container() { if (jQuery("#footnote_references_container").is(":hidden")) { footnote_expand_reference_container(); } else { footnote_collapse_reference_container(); } } function footnote_moveToAnchor(p_str_TargetID) { footnote_expand_reference_container(); var l_obj_Target = jQuery("#" + p_str_TargetID); if(l_obj_Target.length) { jQuery('html, body').animate({ scrollTop: l_obj_Target.offset().top - window.innerHeight/2 }, 1000); } }

The post Revisiting leverage appeared first on The GiveWell Blog.

We’re actively hiring for roles across GiveWell.

Operations

We’re hiring a Director of Operations. The job posting is here.

The Director of Operations is responsible for many domains and manages a team of eight people. A successful candidate will excel at prioritizing the most impactful work, shepherding improvements to completion, and managing the team.

This job is perfect for someone who wants to:

• be part of the leadership team at an organization that’s dedicated to making the world a better place.
• work with colleagues who are passionate about the problems they’re trying to solve.
• have significant personal ownership and responsibility.

We’re looking for someone based in the San Francisco Bay Area, where GiveWell’s office is located. This job has flexible hours and can partly be done remotely.

Outreach

We’re hiring a Head of Growth. The job posting is here.

The Head of Growth will be responsible for leading our efforts to increase the amount of money GiveWell’s recommended charities receive as a result of our recommendation. The Head of Growth will set a strategy to maximize our money moved by identifying, implementing, and testing a variety of growth strategies and will build a team to support these objectives.

We’re looking for a Head of Growth who is excited for the challenge of starting and building our Growth team and aligned with our commitment to honesty and transparency about our, and our recommended organizations’, shortcomings and strengths.

Research

We’re looking for talented people to add to our research team. Some of our most successful analysts are people who followed our work closely prior to joining GiveWell, so if you read our blog, please consider applying!

We’re hiring for three positions:

• Research Analyst and Senior Research Analyst.
• Summer Research Analyst.

Research Analysts and Senior Research Analysts are responsible for all of our research work: reviewing potential top charities and following up with current recommended charities, reviewing the evidence for charitable interventions, building cost-effectiveness models, and evaluating potential Incubation Grants.

Our Summer Research Analyst position is for rising college seniors or graduate students with one year left in their program, and offers the opportunity to work on a variety of research tasks at GiveWell over two to three months.

Research Analysts and Senior Research Analysts do not need to be based in the San Francisco Bay Area. Summer Research Analysts do need to be in the San Francisco Bay Area.

The post GiveWell is hiring! appeared first on The GiveWell Blog.

We’re actively hiring for roles across GiveWell.

Operations

We’re hiring a Director of Operations. The job posting is here.

The Director of Operations is responsible for many domains and manages a team of eight people. A successful candidate will excel at prioritizing the most impactful work, shepherding improvements to completion, and managing the team.

This job is perfect for someone who wants to:

• be part of the leadership team at an organization that’s dedicated to making the world a better place.
• work with colleagues who are passionate about the problems they’re trying to solve.
• have significant personal ownership and responsibility.

We’re looking for someone based in the San Francisco Bay Area, where GiveWell’s office is located. This job has flexible hours and can partly be done remotely.

Outreach

We’re hiring a Head of Growth. The job posting is here.

The Head of Growth will be responsible for leading our efforts to increase the amount of money GiveWell’s recommended charities receive as a result of our recommendation. The Head of Growth will set a strategy to maximize our money moved by identifying, implementing, and testing a variety of growth strategies and will build a team to support these objectives.

We’re looking for a Head of Growth who is excited for the challenge of starting and building our Growth team and aligned with our commitment to honesty and transparency about our, and our recommended organizations’, shortcomings and strengths.

Research

We’re looking for talented people to add to our research team. Some of our most successful analysts are people who followed our work closely prior to joining GiveWell, so if you read our blog, please consider applying!

We’re hiring for three positions:

• Research Analyst and Senior Research Analyst.
• Summer Research Analyst.

Research Analysts and Senior Research Analysts are responsible for all of our research work: reviewing potential top charities and following up with current recommended charities, reviewing the evidence for charitable interventions, building cost-effectiveness models, and evaluating potential Incubation Grants.

Our Summer Research Analyst position is for rising college seniors or graduate students with one year left in their program, and offers the opportunity to work on a variety of research tasks at GiveWell over two to three months.

Research Analysts and Senior Research Analysts do not need to be based in the San Francisco Bay Area. Summer Research Analysts do need to be in the San Francisco Bay Area.

The post GiveWell is hiring! appeared first on The GiveWell Blog.

Summary

• Two of GiveWell’s top charities fight malaria in sub-Saharan Africa.
• GiveWell’s valuations of these charities place some weight on research by Hoyt Bleakley on the impacts of malaria eradication efforts in the American South in the 1920s and in Brazil, Colombia, and Mexico in the 1950s.
• I reviewed the Bleakley study and mostly support its key findings: the campaigns to eradicate malaria from Brazil, Colombia, and Mexico, and perhaps the American South as well, were followed by accelerated income gains for people whose childhood exposure to the disease was reduced. The timing of these events is compatible with the theory that rolling back malaria increased prosperity. Full write-up here.

Introduction

I blogged three weeks ago about having reviewed and reanalyzed Hoyt Bleakley’s study of the effort in the 1910s to rid the American South of hookworm disease (not malaria). That study, published in 2007, seems to show that the children who benefited from the campaign attended school more and went on to earn more as adults.

For GiveWell, Bleakley’s 2010 study is to malaria parasites as his 2007 study is to intestinal worms. Like the 2007 paper, the 2010 one looks back at large-scale, 20th-century eradication campaigns in order to estimate impacts on schooling and adult income. It too produces encouraging results. And it has influenced GiveWell’s recommendations of certain charities—the Against Malaria Foundation and Malaria Consortium’s seasonal malaria chemoprevention program.

Because GiveWell had already invested in replicating and reanalyzing Bleakley (2007), and because the two papers overlap in data and method, I decided to do the same for Bleakley (2010). And here the parallel between the two papers breaks down: having run the evidence through my analytical sieve, my confidence that eradicating malaria boosted income is substantially higher than my confidence that eradicating hookworm did. I’m a bit less sure that it did so in the United States than in Brazil, Colombia, and Mexico; but the Latin American experience is probably more relevant for the places in which our recommended charities work.

This post will walk through the results. For details, see the new working paper. Because my malaria reanalysis shares so much with the hookworm one, I have written this post as if you read the last one. If you haven’t, please do that now.

How the malaria analysis differs from the hookworm one

Having just emphasized the commonality between Bleakley’s hookworm and malaria eradication studies—and my reanalyses thereof—in order to orient you, I should explain how the two differ:

• The hookworm study is set exclusively in the American South, while the malaria study looks at efforts in four countries. In the United States in the 1920s, no doubt inspired by the previous decade’s success against hookworm, the Rockefeller Foundation and the U.S. Public Health Service promoted a large-scale program to drain swamps and spray larvicides, which cut malaria mortality in the South by 60%. Then in the 1950s, with the discovery of DDT, the World Health Organization led a worldwide campaign against the disease. Partly because of data availability, Bleakley (2010) studies the consequences in Brazil, Colombia, and Mexico.1Bleakley (2010) also chose these countries because they had malarial and non-malarial regions, allowing comparisons. See Bleakley (2010), note 6. For sample maps see this. jQuery("#footnote_plugin_tooltip_1").tooltip({ tip: "#footnote_plugin_tooltip_text_1", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });
• Where the hookworm study groups data two ways—first by place of residence to study short-term effects, then by place of birth to study long-term effects—the malaria study does only the latter.
• I pre-registered my analysis plan for the malaria study with the Open Science Framework and hewed to it. While I did not allow the plan to bind my actions, it serves to disclose which analytical tactics I settled on before I touched the data and could know what results they would produce.2Actually we registered a plan for the hookworm study too, but the malaria plan was better informed—and better followed—precisely because it came on the heels of the similar hookworm reanalysis. For brevity, I skipped this theme in my blog post. I did write about it in the hookworm working paper. jQuery("#footnote_plugin_tooltip_2").tooltip({ tip: "#footnote_plugin_tooltip_text_2", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });
• The Bleakley malaria paper appeared in a journal published by the American Economic Association (AEA), which requires its authors to post data and computer code on the AEA website. This aided replication and reanalysis. Unfortunately, as appears to be the norm among AEA journals, the Bleakley (2010) data and code only reproduce the paper’s tables, not the graphs that in this case I see as central.
• For Brazil, Colombia, and Mexico, I mostly relied on that publicly posted data for the crucial information on which regions within a country had the most malaria, rather than trying to construct those variables from old maps and books in Spanish and Portuguese. I also relied on the public data for geographic control variables. I think it can be valuable to go back to primary sources, but for the time being at least, this step looked too time-consuming. I did update and expand the Latin outcome data, on such things as literacy and income, because it is already conveniently digitized in IPUMS International. And I reconstructed all the U.S. data from primary sources, simply by copying what we assembled for the hookworm reanalysis.

Results

In showing you what I found, I’ll follow nearly the same narrative as in my previous post’s section on the “long-term impact on earnings.” To start, here is a key graph from the Bleakley (2010) paper—or really four graphs. In each country’s graph, as with the hookworm graphs, each dot shows the association between historical disease burden in a state (or municipio) and the average income in adulthood of people born there in a given year. In all but Colombia, the leftmost dots line up with the negative range on the vertical axis, meaning that, initially, coming from a historically malarial area stunted one’s income. For example, some of the early U.S. dots are around –0.1 on the vertical axis, which means that being native to swampy Mississippi instead of arid Wyoming cut one’s adult earnings by about 10%.3For cross-country comparability, Bleakley (2010) normalizes the malaria mortality and ecology indexes so that the 5th- and 95th-percentile geographic units—Wyoming and Mississippi in the U.S. case—score 0 and 1. Income proxies are taken in logs. jQuery("#footnote_plugin_tooltip_3").tooltip({ tip: "#footnote_plugin_tooltip_text_3", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); The dots later rise, suggesting that the liability of coming from malarial areas faded, and even reversed. In Colombia, the dots start around zero but also then rise.

As in the hookworm study, here, Bleakley (2010) superimposes on the dots the step-like contour representing how malaria eradication is expected to play out in the data. The steps reach their full height when the campaigns are taken to have started—1920 in the United States and 1957 in the Latin nations. All babies born after these points were alike in that they grew up fully in the post–eradication campaign world. The step contours begin their rises 18 years earlier, when the first babies were born who would benefit from eradication at least a bit by their 18th birthdays.4These graphs incorporate all of Bleakley’s control variables. In my hookworm post, I began both results sections with “basic” graphs that did not include all the controls, imitating Bleakley (2007). In contrast, all the Bleakley (2010) graphs incorporate full controls. So I do the same. jQuery("#footnote_plugin_tooltip_4").tooltip({ tip: "#footnote_plugin_tooltip_text_4", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

Next is my closest replication of the key Bleakley (2010) graphs. These use Bleakley’s data, as posted, but not Bleakley’s computer code, since that was not posted:

The next version adds the latest rounds of census data from the Latin nations and the newer, larger samples from old census rounds for the United States. It also redefines childhood as lasting 21 instead of 18 years, because I discovered that the Bleakley (2010) code uses 18 but the text uses 21. That budges the first dashed lines back by three years:

I avoided superimposing step contours on these data points because I worried that it would trick the brain into thinking that the contours fit the data better than they do. But whether the step contour fits the plots above is exactly what you should ask yourself now. Does it seem as if the dots rise, or rise more, between each pair of vertical, dashed lines? I could see the answer being “yes” for all but Mexico. And that could be a fingerprint of malaria eradication.

I ask that question more formally in the next quartet, fitting line segments to successive ranges of the data. The dots in the four graphs are the same as above, but I’ve taken away the grey confidence intervals for readability. The p values in the lower-left of each pane speak to whether any upward or downward bends at the allowed kink points are statistically significant, i.e., hard to ascribe to chance alone. Where the p values are low—and they mostly are, even in Mexico—they favor the Bleakley (2010) reading that rolling back malaria raised incomes.

In Brazil, Colombia, and Mexico, this statistical test is fairly confident that red lines bend upward at the first kinks (p = 0.00 for Brazil and Colombia and 0.07 for Mexico). That is: in high-malaria areas, relative to low-malaria areas, as the first babies were born who could benefit in childhood from eradication, future incomes rose. The test is less confident for the United States, where the first allowed kink, in 1899, gets a high-ish p value of 0.39. However, the U.S. trend clearly bends upward—just earlier than predicted by the Bleakley (2010) theory. That might mean that the Bleakley (2010) theory is slightly wrong: maybe when it came to impacts on future earnings, malaria exposure continued to matter into one’s twenties, at least in the United States 100 years ago. Then, people born in the South even a bit before 1899 (the date of the first U.S. kink point) would have benefited from the eventual campaign against malaria; and that first kink should be moved to the left, where it would match the data better and produce a lower p value. Or perhaps that high p value of 0.39 signifies that the Bleakley (2010) model is completely wrong for the United States, and that forces other than malaria eradication drove the South’s catch-up on income.

Now, in addition to the four measures of income studied above–one for each country—the Bleakley (2010) paper looks at eight other outcomes. Six are literacy and years of schooling completed, tracked separately in Brazil, Colombia, and Mexico. In addition, there is, for Brazil, earned income—as distinct from total income (“earned” meaning earned through work). And there is, for the United States, Duncan’s Socioeconomic Index (SEI), which blends the occupational income score, explained in my last post, with information about a person’s education level. Your Duncan’s SEI is highest if you hold what is typically a high-paying job (as with the occupational income score) and you have a lot of education.

The first public version of the Bleakley study makes graphs for the additional eight outcomes too. But the final, journal-published version drops them, perhaps to save space. Since for me, the graphs are so central, I generated my own graphs for the other eight outcomes:

These figures hand us a mixed bag. In the United States, the trend on Duncan’s index appears to bend as predicted at the first allowed kink (p = 0.04) but not the second. Seemingly, relative income gains continued in the South well after malaria eradication could cause them. In Brazil, while relative progress on earned income slows when expected (second kink, p = 0.04), it does not appear to accelerate when expected (first kink), perhaps owing to small samples in the early years. In none of the Latin countries does relative progress on adult literacy or years of schooling slow with much statistical significance at the expected time (second kink points in bottom six graphs). The trend bends in all six at the first kink point, and with statistical significance—but the wrong way in Mexico.

In fact, the mixed bag partly corroborates Bleakley (2010), which also questions whether rolling back malaria increased schooling. The new results depart from Bleakley (2010) in also questioning the benefit for literacy. And they cast some doubt on the income impact in the United States. In both the U.S. plots—in the upper-left of the last two sets of graphs above—it’s clear that the income gap between the South and the rest narrowed over many decades. It’s less clear that it did so with a rhythm attributable to the malaria eradication effort of the 1920s.

Conclusion

For me, this reanalysis triggers a modest update to my understanding of the impacts of malaria prevention. With regard to adult income in Latin America, and perhaps the United States, the Bleakley (2010) theory withstands reexamination. It holds up less well for literacy, but this is not very surprising given that Bleakley (2010) also did not find clear impacts on schooling.

I wouldn’t say that my confirmation proves that malaria eradication campaigns in the Americas boosted income in the way that a large-scale randomized study might. But then neither, if you read him closely, does Bleakley. Rather, the evidence “indicates” impact. The theory that malaria eradication in the Americas increased earnings fits pretty well to the data we have. And that is probably about as much certainty as we can expect from this historical analysis.

Much of the data and code for this study are here (2 GB). Because of IPUMS licensing limitations, the download leaves out the census data for Brazil, Colombia, and Mexico. The included “read me” file explains how to obtain this data. The full write-up is here.

Notes   [ + ]

1. ↑ Bleakley (2010) also chose these countries because they had malarial and non-malarial regions, allowing comparisons. See Bleakley (2010), note 6. For sample maps see this. 2. ↑ Actually we registered a plan for the hookworm study too, but the malaria plan was better informed—and better followed—precisely because it came on the heels of the similar hookworm reanalysis. For brevity, I skipped this theme in my blog post. I did write about it in the hookworm working paper. 3. ↑ For cross-country comparability, Bleakley (2010) normalizes the malaria mortality and ecology indexes so that the 5th- and 95th-percentile geographic units—Wyoming and Mississippi in the U.S. case—score 0 and 1. Income proxies are taken in logs. 4. ↑ These graphs incorporate all of Bleakley’s control variables. In my hookworm post, I began both results sections with “basic” graphs that did not include all the controls, imitating Bleakley (2007). In contrast, all the Bleakley (2010) graphs incorporate full controls. So I do the same. function footnote_expand_reference_container() { jQuery("#footnote_references_container").show(); jQuery("#footnote_reference_container_collapse_button").text("-"); } function footnote_collapse_reference_container() { jQuery("#footnote_references_container").hide(); jQuery("#footnote_reference_container_collapse_button").text("+"); } function footnote_expand_collapse_reference_container() { if (jQuery("#footnote_references_container").is(":hidden")) { footnote_expand_reference_container(); } else { footnote_collapse_reference_container(); } } function footnote_moveToAnchor(p_str_TargetID) { footnote_expand_reference_container(); var l_obj_Target = jQuery("#" + p_str_TargetID); if(l_obj_Target.length) { jQuery('html, body').animate({ scrollTop: l_obj_Target.offset().top - window.innerHeight/2 }, 1000); } }

The post Revisiting the evidence on malaria eradication in the Americas appeared first on The GiveWell Blog.

Summary

• Two of GiveWell’s top charities fight malaria in sub-Saharan Africa.
• GiveWell’s valuations of these charities place some weight on research by Hoyt Bleakley on the impacts of malaria eradication efforts in the American South in the 1920s and in Brazil, Colombia, and Mexico in the 1950s.
• I reviewed the Bleakley study and mostly support its key findings: the campaigns to eradicate malaria from Brazil, Colombia, and Mexico, and perhaps the American South as well, were followed by accelerated income gains for people whose childhood exposure to the disease was reduced. The timing of these events is compatible with the theory that rolling back malaria increased prosperity. Full write-up here.

Introduction

I blogged three weeks ago about having reviewed and reanalyzed Hoyt Bleakley’s study of the effort in the 1910s to rid the American South of hookworm disease (not malaria). That study, published in 2007, seems to show that the children who benefited from the campaign attended school more and went on to earn more as adults.

For GiveWell, Bleakley’s 2010 study is to malaria parasites as his 2007 study is to intestinal worms. Like the 2007 paper, the 2010 one looks back at large-scale, 20th-century eradication campaigns in order to estimate impacts on schooling and adult income. It too produces encouraging results. And it has influenced GiveWell’s recommendations of certain charities—the Against Malaria Foundation and Malaria Consortium’s seasonal malaria chemoprevention program.

Because GiveWell had already invested in replicating and reanalyzing Bleakley (2007), and because the two papers overlap in data and method, I decided to do the same for Bleakley (2010). And here the parallel between the two papers breaks down: having run the evidence through my analytical sieve, my confidence that eradicating malaria boosted income is substantially higher than my confidence that eradicating hookworm did. I’m a bit less sure that it did so in the United States than in Brazil, Colombia, and Mexico; but the Latin American experience is probably more relevant for the places in which our recommended charities work.

This post will walk through the results. For details, see the new working paper. Because my malaria reanalysis shares so much with the hookworm one, I have written this post as if you read the last one. If you haven’t, please do that now.

How the malaria analysis differs from the hookworm one

Having just emphasized the commonality between Bleakley’s hookworm and malaria eradication studies—and my reanalyses thereof—in order to orient you, I should explain how the two differ:

• The hookworm study is set exclusively in the American South, while the malaria study looks at efforts in four countries. In the United States in the 1920s, no doubt inspired by the previous decade’s success against hookworm, the Rockefeller Foundation and the U.S. Public Health Service promoted a large-scale program to drain swamps and spray larvicides, which cut malaria mortality in the South by 60%. Then in the 1950s, with the discovery of DDT, the World Health Organization led a worldwide campaign against the disease. Partly because of data availability, Bleakley (2010) studies the consequences in Brazil, Colombia, and Mexico.1Bleakley (2010) also chose these countries because they had malarial and non-malarial regions, allowing comparisons. See Bleakley (2010), note 6. For sample maps see this. jQuery("#footnote_plugin_tooltip_1").tooltip({ tip: "#footnote_plugin_tooltip_text_1", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });
• Where the hookworm study groups data two ways—first by place of residence to study short-term effects, then by place of birth to study long-term effects—the malaria study does only the latter.
• I pre-registered my analysis plan for the malaria study with the Open Science Framework and hewed to it. While I did not allow the plan to bind my actions, it serves to disclose which analytical tactics I settled on before I touched the data and could know what results they would produce.2Actually we registered a plan for the hookworm study too, but the malaria plan was better informed—and better followed—precisely because it came on the heels of the similar hookworm reanalysis. For brevity, I skipped this theme in my blog post. I did write about it in the hookworm working paper. jQuery("#footnote_plugin_tooltip_2").tooltip({ tip: "#footnote_plugin_tooltip_text_2", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });
• The Bleakley malaria paper appeared in a journal published by the American Economic Association (AEA), which requires its authors to post data and computer code on the AEA website. This aided replication and reanalysis. Unfortunately, as appears to be the norm among AEA journals, the Bleakley (2010) data and code only reproduce the paper’s tables, not the graphs that in this case I see as central.
• For Brazil, Colombia, and Mexico, I mostly relied on that publicly posted data for the crucial information on which regions within a country had the most malaria, rather than trying to construct those variables from old maps and books in Spanish and Portuguese. I also relied on the public data for geographic control variables. I think it can be valuable to go back to primary sources, but for the time being at least, this step looked too time-consuming. I did update and expand the Latin outcome data, on such things as literacy and income, because it is already conveniently digitized in IPUMS International. And I reconstructed all the U.S. data from primary sources, simply by copying what we assembled for the hookworm reanalysis.

Results

In showing you what I found, I’ll follow nearly the same narrative as in my previous post’s section on the “long-term impact on earnings.” To start, here is a key graph from the Bleakley (2010) paper—or really four graphs. In each country’s graph, as with the hookworm graphs, each dot shows the association between historical disease burden in a state (or municipio) and the average income in adulthood of people born there in a given year. In all but Colombia, the leftmost dots line up with the negative range on the vertical axis, meaning that, initially, coming from a historically malarial area stunted one’s income. For example, some of the early U.S. dots are around –0.1 on the vertical axis, which means that being native to swampy Mississippi instead of arid Wyoming cut one’s adult earnings by about 10%.3For cross-country comparability, Bleakley (2010) normalizes the malaria mortality and ecology indexes so that the 5th- and 95th-percentile geographic units—Wyoming and Mississippi in the U.S. case—score 0 and 1. Income proxies are taken in logs. jQuery("#footnote_plugin_tooltip_3").tooltip({ tip: "#footnote_plugin_tooltip_text_3", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); The dots later rise, suggesting that the liability of coming from malarial areas faded, and even reversed. In Colombia, the dots start around zero but also then rise.

As in the hookworm study, here, Bleakley (2010) superimposes on the dots the step-like contour representing how malaria eradication is expected to play out in the data. The steps reach their full height when the campaigns are taken to have started—1920 in the United States and 1957 in the Latin nations. All babies born after these points were alike in that they grew up fully in the post–eradication campaign world. The step contours begin their rises 18 years earlier, when the first babies were born who would benefit from eradication at least a bit by their 18th birthdays.4These graphs incorporate all of Bleakley’s control variables. In my hookworm post, I began both results sections with “basic” graphs that did not include all the controls, imitating Bleakley (2007). In contrast, all the Bleakley (2010) graphs incorporate full controls. So I do the same. jQuery("#footnote_plugin_tooltip_4").tooltip({ tip: "#footnote_plugin_tooltip_text_4", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

Next is my closest replication of the key Bleakley (2010) graphs. These use Bleakley’s data, as posted, but not Bleakley’s computer code, since that was not posted:

The next version adds the latest rounds of census data from the Latin nations and the newer, larger samples from old census rounds for the United States. It also redefines childhood as lasting 21 instead of 18 years, because I discovered that the Bleakley (2010) code uses 18 but the text uses 21. That budges the first dashed lines back by three years:

I avoided superimposing step contours on these data points because I worried that it would trick the brain into thinking that the contours fit the data better than they do. But whether the step contour fits the plots above is exactly what you should ask yourself now. Does it seem as if the dots rise, or rise more, between each pair of vertical, dashed lines? I could see the answer being “yes” for all but Mexico. And that could be a fingerprint of malaria eradication.

I ask that question more formally in the next quartet, fitting line segments to successive ranges of the data. The dots in the four graphs are the same as above, but I’ve taken away the grey confidence intervals for readability. The p values in the lower-left of each pane speak to whether any upward or downward bends at the allowed kink points are statistically significant, i.e., hard to ascribe to chance alone. Where the p values are low—and they mostly are, even in Mexico—they favor the Bleakley (2010) reading that rolling back malaria raised incomes.

In Brazil, Colombia, and Mexico, this statistical test is fairly confident that red lines bend upward at the first kinks (p = 0.00 for Brazil and Colombia and 0.07 for Mexico). That is: in high-malaria areas, relative to low-malaria areas, as the first babies were born who could benefit in childhood from eradication, future incomes rose. The test is less confident for the United States, where the first allowed kink, in 1899, gets a high-ish p value of 0.39. However, the U.S. trend clearly bends upward—just earlier than predicted by the Bleakley (2010) theory. That might mean that the Bleakley (2010) theory is slightly wrong: maybe when it came to impacts on future earnings, malaria exposure continued to matter into one’s twenties, at least in the United States 100 years ago. Then, people born in the South even a bit before 1899 (the date of the first U.S. kink point) would have benefited from the eventual campaign against malaria; and that first kink should be moved to the left, where it would match the data better and produce a lower p value. Or perhaps that high p value of 0.39 signifies that the Bleakley (2010) model is completely wrong for the United States, and that forces other than malaria eradication drove the South’s catch-up on income.

Now, in addition to the four measures of income studied above–one for each country—the Bleakley (2010) paper looks at eight other outcomes. Six are literacy and years of schooling completed, tracked separately in Brazil, Colombia, and Mexico. In addition, there is, for Brazil, earned income—as distinct from total income (“earned” meaning earned through work). And there is, for the United States, Duncan’s Socioeconomic Index (SEI), which blends the occupational income score, explained in my last post, with information about a person’s education level. Your Duncan’s SEI is highest if you hold what is typically a high-paying job (as with the occupational income score) and you have a lot of education.

The first public version of the Bleakley study makes graphs for the additional eight outcomes too. But the final, journal-published version drops them, perhaps to save space. Since for me, the graphs are so central, I generated my own graphs for the other eight outcomes:

These figures hand us a mixed bag. In the United States, the trend on Duncan’s index appears to bend as predicted at the first allowed kink (p = 0.04) but not the second. Seemingly, relative income gains continued in the South well after malaria eradication could cause them. In Brazil, while relative progress on earned income slows when expected (second kink, p = 0.04), it does not appear to accelerate when expected (first kink), perhaps owing to small samples in the early years. In none of the Latin countries does relative progress on adult literacy or years of schooling slow with much statistical significance at the expected time (second kink points in bottom six graphs). The trend bends in all six at the first kink point, and with statistical significance—but the wrong way in Mexico.

In fact, the mixed bag partly corroborates Bleakley (2010), which also questions whether rolling back malaria increased schooling. The new results depart from Bleakley (2010) in also questioning the benefit for literacy. And they cast some doubt on the income impact in the United States. In both the U.S. plots—in the upper-left of the last two sets of graphs above—it’s clear that the income gap between the South and the rest narrowed over many decades. It’s less clear that it did so with a rhythm attributable to the malaria eradication effort of the 1920s.

Conclusion

For me, this reanalysis triggers a modest update to my understanding of the impacts of malaria prevention. With regard to adult income in Latin America, and perhaps the United States, the Bleakley (2010) theory withstands reexamination. It holds up less well for literacy, but this is not very surprising given that Bleakley (2010) also did not find clear impacts on schooling.

I wouldn’t say that my confirmation proves that malaria eradication campaigns in the Americas boosted income in the way that a large-scale randomized study might. But then neither, if you read him closely, does Bleakley. Rather, the evidence “indicates” impact. The theory that malaria eradication in the Americas increased earnings fits pretty well to the data we have. And that is probably about as much certainty as we can expect from this historical analysis.

Much of the data and code for this study are here (2 GB). Because of IPUMS licensing limitations, the download leaves out the census data for Brazil, Colombia, and Mexico. The included “read me” file explains how to obtain this data. The full write-up is here.

Notes   [ + ]

1. ↑ Bleakley (2010) also chose these countries because they had malarial and non-malarial regions, allowing comparisons. See Bleakley (2010), note 6. For sample maps see this. 2. ↑ Actually we registered a plan for the hookworm study too, but the malaria plan was better informed—and better followed—precisely because it came on the heels of the similar hookworm reanalysis. For brevity, I skipped this theme in my blog post. I did write about it in the hookworm working paper. 3. ↑ For cross-country comparability, Bleakley (2010) normalizes the malaria mortality and ecology indexes so that the 5th- and 95th-percentile geographic units—Wyoming and Mississippi in the U.S. case—score 0 and 1. Income proxies are taken in logs. 4. ↑ These graphs incorporate all of Bleakley’s control variables. In my hookworm post, I began both results sections with “basic” graphs that did not include all the controls, imitating Bleakley (2007). In contrast, all the Bleakley (2010) graphs incorporate full controls. So I do the same. function footnote_expand_reference_container() { jQuery("#footnote_references_container").show(); jQuery("#footnote_reference_container_collapse_button").text("-"); } function footnote_collapse_reference_container() { jQuery("#footnote_references_container").hide(); jQuery("#footnote_reference_container_collapse_button").text("+"); } function footnote_expand_collapse_reference_container() { if (jQuery("#footnote_references_container").is(":hidden")) { footnote_expand_reference_container(); } else { footnote_collapse_reference_container(); } } function footnote_moveToAnchor(p_str_TargetID) { footnote_expand_reference_container(); var l_obj_Target = jQuery("#" + p_str_TargetID); if(l_obj_Target.length) { jQuery('html, body').animate({ scrollTop: l_obj_Target.offset().top - window.innerHeight/2 }, 1000); } }

The post Revisiting the evidence on malaria eradication in the Americas appeared first on The GiveWell Blog.

Summary

• Two of GiveWell’s top charities fight malaria in sub-Saharan Africa.
• GiveWell’s valuations of these charities place some weight on research by Hoyt Bleakley on the impacts of malaria eradication efforts in the American South in the 1920s and in Brazil, Colombia, and Mexico in the 1950s.
• I reviewed the Bleakley study and mostly support its key findings: the campaigns to eradicate malaria from Brazil, Colombia, and Mexico, and perhaps the American South as well, were followed by accelerated income gains for people whose childhood exposure to the disease was reduced. The timing of these events is compatible with the theory that rolling back malaria increased prosperity. Full write-up here.

Introduction

I blogged three weeks ago about having reviewed and reanalyzed Hoyt Bleakley’s study of the effort in the 1910s to rid the American South of hookworm disease (not malaria). That study, published in 2007, seems to show that the children who benefited from the campaign attended school more and went on to earn more as adults.

For GiveWell, Bleakley’s 2010 study is to malaria parasites as his 2007 study is to intestinal worms. Like the 2007 paper, the 2010 one looks back at large-scale, 20th-century eradication campaigns in order to estimate impacts on schooling and adult income. It too produces encouraging results. And it has influenced GiveWell’s recommendations of certain charities—the Against Malaria Foundation and Malaria Consortium’s seasonal malaria chemoprevention program.

Because GiveWell had already invested in replicating and reanalyzing Bleakley (2007), and because the two papers overlap in data and method, I decided to do the same for Bleakley (2010). And here the parallel between the two papers breaks down: having run the evidence through my analytical sieve, my confidence that eradicating malaria boosted income is substantially higher than my confidence that eradicating hookworm did. I’m a bit less sure that it did so in the United States than in Brazil, Colombia, and Mexico; but the Latin American experience is probably more relevant for the places in which our recommended charities work.

This post will walk through the results. For details, see the new working paper. Because my malaria reanalysis shares so much with the hookworm one, I have written this post as if you read the last one. If you haven’t, please do that now.

How the malaria analysis differs from the hookworm one

Having just emphasized the commonality between Bleakley’s hookworm and malaria eradication studies—and my reanalyses thereof—in order to orient you, I should explain how the two differ:

• The hookworm study is set exclusively in the American South, while the malaria study looks at efforts in four countries. In the United States in the 1920s, no doubt inspired by the previous decade’s success against hookworm, the Rockefeller Foundation and the U.S. Public Health Service promoted a large-scale program to drain swamps and spray larvicides, which cut malaria mortality in the South by 60%. Then in the 1950s, with the discovery of DDT, the World Health Organization led a worldwide campaign against the disease. Partly because of data availability, Bleakley (2010) studies the consequences in Brazil, Colombia, and Mexico.1Bleakley (2010) also chose these countries because they had malarial and non-malarial regions, allowing comparisons. See Bleakley (2010), note 6. For sample maps see this. jQuery("#footnote_plugin_tooltip_1").tooltip({ tip: "#footnote_plugin_tooltip_text_1", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });
• Where the hookworm study groups data two ways—first by place of residence to study short-term effects, then by place of birth to study long-term effects—the malaria study does only the latter.
• I pre-registered my analysis plan for the malaria study with the Open Science Framework and hewed to it. While I did not allow the plan to bind my actions, it serves to disclose which analytical tactics I settled on before I touched the data and could know what results they would produce.2Actually we registered a plan for the hookworm study too, but the malaria plan was better informed—and better followed—precisely because it came on the heels of the similar hookworm reanalysis. For brevity, I skipped this theme in my blog post. I did write about it in the hookworm working paper. jQuery("#footnote_plugin_tooltip_2").tooltip({ tip: "#footnote_plugin_tooltip_text_2", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });
• The Bleakley malaria paper appeared in a journal published by the American Economic Association (AEA), which requires its authors to post data and computer code on the AEA website. This aided replication and reanalysis. Unfortunately, as appears to be the norm among AEA journals, the Bleakley (2010) data and code only reproduce the paper’s tables, not the graphs that in this case I see as central.
• For Brazil, Colombia, and Mexico, I mostly relied on that publicly posted data for the crucial information on which regions within a country had the most malaria, rather than trying to construct those variables from old maps and books in Spanish and Portuguese. I also relied on the public data for geographic control variables. I think it can be valuable to go back to primary sources, but for the time being at least, this step looked too time-consuming. I did update and expand the Latin outcome data, on such things as literacy and income, because it is already conveniently digitized in IPUMS International. And I reconstructed all the U.S. data from primary sources, simply by copying what we assembled for the hookworm reanalysis.

Results

In showing you what I found, I’ll follow nearly the same narrative as in my previous post’s section on the “long-term impact on earnings.” To start, here is a key graph from the Bleakley (2010) paper—or really four graphs. In each country’s graph, as with the hookworm graphs, each dot shows the association between historical disease burden in a state (or municipio) and the average income in adulthood of people born there in a given year. In all but Colombia, the leftmost dots line up with the negative range on the vertical axis, meaning that, initially, coming from a historically malarial area stunted one’s income. For example, some of the early U.S. dots are around –0.1 on the vertical axis, which means that being native to swampy Mississippi instead of arid Wyoming cut one’s adult earnings by about 10%.3For cross-country comparability, Bleakley (2010) normalizes the malaria mortality and ecology indexes so that the 5th- and 95th-percentile geographic units—Wyoming and Mississippi in the U.S. case—score 0 and 1. Income proxies are taken in logs. jQuery("#footnote_plugin_tooltip_3").tooltip({ tip: "#footnote_plugin_tooltip_text_3", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); The dots later rise, suggesting that the liability of coming from malarial areas faded, and even reversed. In Colombia, the dots start around zero but also then rise.

As in the hookworm study, here, Bleakley (2010) superimposes on the dots the step-like contour representing how malaria eradication is expected to play out in the data. The steps reach their full height when the campaigns are taken to have started—1920 in the United States and 1957 in the Latin nations. All babies born after these points were alike in that they grew up fully in the post–eradication campaign world. The step contours begin their rises 18 years earlier, when the first babies were born who would benefit from eradication at least a bit by their 18th birthdays.4These graphs incorporate all of Bleakley’s control variables. In my hookworm post, I began both results sections with “basic” graphs that did not include all the controls, imitating Bleakley (2007). In contrast, all the Bleakley (2010) graphs incorporate full controls. So I do the same. jQuery("#footnote_plugin_tooltip_4").tooltip({ tip: "#footnote_plugin_tooltip_text_4", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

Next is my closest replication of the key Bleakley (2010) graphs. These use Bleakley’s data, as posted, but not Bleakley’s computer code, since that was not posted:

The next version adds the latest rounds of census data from the Latin nations and the newer, larger samples from old census rounds for the United States. It also redefines childhood as lasting 21 instead of 18 years, because I discovered that the Bleakley (2010) code uses 18 but the text uses 21. That budges the first dashed lines back by three years:

I avoided superimposing step contours on these data points because I worried that it would trick the brain into thinking that the contours fit the data better than they do. But whether the step contour fits the plots above is exactly what you should ask yourself now. Does it seem as if the dots rise, or rise more, between each pair of vertical, dashed lines? I could see the answer being “yes” for all but Mexico. And that could be a fingerprint of malaria eradication.

I ask that question more formally in the next quartet, fitting line segments to successive ranges of the data. The dots in the four graphs are the same as above, but I’ve taken away the grey confidence intervals for readability. The p values in the lower-left of each pane speak to whether any upward or downward bends at the allowed kink points are statistically significant, i.e., hard to ascribe to chance alone. Where the p values are low—and they mostly are, even in Mexico—they favor the Bleakley (2010) reading that rolling back malaria raised incomes.

In Brazil, Colombia, and Mexico, this statistical test is fairly confident that red lines bend upward at the first kinks (p = 0.00 for Brazil and Colombia and 0.07 for Mexico). That is: in high-malaria areas, relative to low-malaria areas, as the first babies were born who could benefit in childhood from eradication, future incomes rose. The test is less confident for the United States, where the first allowed kink, in 1899, gets a high-ish p value of 0.39. However, the U.S. trend clearly bends upward—just earlier than predicted by the Bleakley (2010) theory. That might mean that the Bleakley (2010) theory is slightly wrong: maybe when it came to impacts on future earnings, malaria exposure continued to matter into one’s twenties, at least in the United States 100 years ago. Then, people born in the South even a bit before 1899 (the date of the first U.S. kink point) would have benefited from the eventual campaign against malaria; and that first kink should be moved to the left, where it would match the data better and produce a lower p value. Or perhaps that high p value of 0.39 signifies that the Bleakley (2010) model is completely wrong for the United States, and that forces other than malaria eradication drove the South’s catch-up on income.

Now, in addition to the four measures of income studied above–one for each country—the Bleakley (2010) paper looks at eight other outcomes. Six are literacy and years of schooling completed, tracked separately in Brazil, Colombia, and Mexico. In addition, there is, for Brazil, earned income—as distinct from total income (“earned” meaning earned through work). And there is, for the United States, Duncan’s Socioeconomic Index (SEI), which blends the occupational income score, explained in my last post, with information about a person’s education level. Your Duncan’s SEI is highest if you hold what is typically a high-paying job (as with the occupational income score) and you have a lot of education.

The first public version of the Bleakley study makes graphs for the additional eight outcomes too. But the final, journal-published version drops them, perhaps to save space. Since for me, the graphs are so central, I generated my own graphs for the other eight outcomes:

These figures hand us a mixed bag. In the United States, the trend on Duncan’s index appears to bend as predicted at the first allowed kink (p = 0.04) but not the second. Seemingly, relative income gains continued in the South well after malaria eradication could cause them. In Brazil, while relative progress on earned income slows when expected (second kink, p = 0.04), it does not appear to accelerate when expected (first kink), perhaps owing to small samples in the early years. In none of the Latin countries does relative progress on adult literacy or years of schooling slow with much statistical significance at the expected time (second kink points in bottom six graphs). The trend bends in all six at the first kink point, and with statistical significance—but the wrong way in Mexico.

In fact, the mixed bag partly corroborates Bleakley (2010), which also questions whether rolling back malaria increased schooling. The new results depart from Bleakley (2010) in also questioning the benefit for literacy. And they cast some doubt on the income impact in the United States. In both the U.S. plots—in the upper-left of the last two sets of graphs above—it’s clear that the income gap between the South and the rest narrowed over many decades. It’s less clear that it did so with a rhythm attributable to the malaria eradication effort of the 1920s.

Conclusion

For me, this reanalysis triggers a modest update to my understanding of the impacts of malaria prevention. With regard to adult income in Latin America, and perhaps the United States, the Bleakley (2010) theory withstands reexamination. It holds up less well for literacy, but this is not very surprising given that Bleakley (2010) also did not find clear impacts on schooling.

I wouldn’t say that my confirmation proves that malaria eradication campaigns in the Americas boosted income in the way that a large-scale randomized study might. But then neither, if you read him closely, does Bleakley. Rather, the evidence “indicates” impact. The theory that malaria eradication in the Americas increased earnings fits pretty well to the data we have. And that is probably about as much certainty as we can expect from this historical analysis.

Much of the data and code for this study are here (2 GB). Because of IPUMS licensing limitations, the download leaves out the census data for Brazil, Colombia, and Mexico. The included “read me” file explains how to obtain this data. The full write-up is here.

Notes   [ + ]

1. ↑ Bleakley (2010) also chose these countries because they had malarial and non-malarial regions, allowing comparisons. See Bleakley (2010), note 6. For sample maps see this. 2. ↑ Actually we registered a plan for the hookworm study too, but the malaria plan was better informed—and better followed—precisely because it came on the heels of the similar hookworm reanalysis. For brevity, I skipped this theme in my blog post. I did write about it in the hookworm working paper. 3. ↑ For cross-country comparability, Bleakley (2010) normalizes the malaria mortality and ecology indexes so that the 5th- and 95th-percentile geographic units—Wyoming and Mississippi in the U.S. case—score 0 and 1. Income proxies are taken in logs. 4. ↑ These graphs incorporate all of Bleakley’s control variables. In my hookworm post, I began both results sections with “basic” graphs that did not include all the controls, imitating Bleakley (2007). In contrast, all the Bleakley (2010) graphs incorporate full controls. So I do the same. function footnote_expand_reference_container() { jQuery("#footnote_references_container").show(); jQuery("#footnote_reference_container_collapse_button").text("-"); } function footnote_collapse_reference_container() { jQuery("#footnote_references_container").hide(); jQuery("#footnote_reference_container_collapse_button").text("+"); } function footnote_expand_collapse_reference_container() { if (jQuery("#footnote_references_container").is(":hidden")) { footnote_expand_reference_container(); } else { footnote_collapse_reference_container(); } } function footnote_moveToAnchor(p_str_TargetID) { footnote_expand_reference_container(); var l_obj_Target = jQuery("#" + p_str_TargetID); if(l_obj_Target.length) { jQuery('html, body').animate({ scrollTop: l_obj_Target.offset().top - window.innerHeight/2 }, 1000); } }

The post Revisiting the evidence on malaria eradication in the Americas appeared first on The GiveWell Blog.

One of our two new top charities this year is Helen Keller International (HKI)’s vitamin A supplementation program. We named HKI’s vitamin A supplementation program a top charity this year because:

• There is strong evidence from many randomized controlled trials of vitamin A supplementation that the program leads to substantial reductions in child deaths.
• HKI-supported vitamin A supplementation programs are inexpensive (we estimate around $0.75 in total costs per supplement delivered) and highly cost-effective at preventing child deaths in countries where HKI plans to work using GiveWell-directed funds.
• HKI is transparent—it has shared significant, detailed information about its programs with us, including the results and methodology of monitoring surveys HKI conducted to determine whether its vitamin A supplementation programs reach a large proportion of targeted children.
• HKI has a funding gap—we believe it is highly likely that its vitamin A supplementation programs will be constrained by funding next year.

HKI’s vitamin A supplementation program is an exceptional giving opportunity, but as with the case for donating to any of our other top charities, not a “sure thing.”

I’m the Research Analyst who has led our work on HKI this year. In this post, I discuss some key questions about the impact of Helen Keller International’s vitamin A supplementation program and what we’ve learned so far. I also discuss GiveWell’s plans for learning more about these issues in the future.

In short:

• Is vitamin A deficiency still a major concern? Our best guess is that vitamin A deficiency is considerably less common today where HKI works than it was among children who participated in past trials of vitamin A supplementation, but not so rare that vitamin A supplementation would not be cost-effective. We are quite uncertain about our estimate of the prevalence of vitamin A deficiency where HKI works because little high-quality, up-to-date data on vitamin A deficiency is available. We plan to consider funding new surveys of vitamin A deficiency to improve our understanding of the effectiveness of HKI’s programs.
• Have improvements in health conditions over time reduced the need for vitamin A supplementation? Child mortality rates remain quite high in areas where HKI plans to use GiveWell-directed funding for vitamin A supplementation programs. We think it’s unlikely that health conditions in these countries have improved enough for vitamin A supplementation to no longer be effective.
• How strong is HKI’s track record of supporting fixed-point vitamin A supplement distributions? HKI expects to primarily support fixed-point vitamin A supplement distributions (rather than door-to-door campaigns) going forward. Results from monitoring surveys have found that, on average, HKI’s fixed-point programs have not reached as high a proportion of targeted populations as its door-to-door programs, but these monitoring surveys may not have been fully representative of HKI’s programs overall. Our best guess is that future fixed-point programs will achieve moderate to high coverage.

Is vitamin A deficiency still a major concern?

Vitamin A deficiency, a condition resulting from chronic low vitamin A intake, can cause loss of vision and increased severity of infections. If vitamin A deficiency is less common today than it was among participants in trials of vitamin A supplementation, today’s programs may prevent fewer deaths than the evidence from the trials suggests.

We estimate that the prevalence of vitamin A deficiency was high (around 60%) in the populations studied in trials included in the Cochrane Collaboration review of vitamin A supplementation programs for preschool-aged children, Imdad et al. 2017.1See the “Imdad 2017 – VAD prevalence estimates” sheet here for details. jQuery("#footnote_plugin_tooltip_1").tooltip({ tip: "#footnote_plugin_tooltip_text_1", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

The map below, from Our World in Data, presents the World Health Organization (WHO)’s most recent estimates of the prevalence of vitamin A deficiency among preschool-aged children by country, covering the period from 1995 to 2005. WHO categorizes prevalences of vitamin A deficiency among preschool-aged children of 20% or above as a severe public health problem.2WHO Global prevalence of vitamin A deficiency in populations at risk 2009, Pg 8, Table 5. jQuery("#footnote_plugin_tooltip_2").tooltip({ tip: "#footnote_plugin_tooltip_text_2", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

Since WHO’s most recent estimates are now considerably out-of-date, we decided to investigate a variety of additional sources in order to create best-guess estimates of rates of vitamin A deficiency today in countries in sub-Saharan Africa where HKI works.

We learned that there is very little useful, up-to-date data on vitamin A deficiency in countries in sub-Saharan Africa. In many countries, the most recent surveys of vitamin A deficiency were completed ten or more years ago. Many governments have also recently mandated the fortification of vegetable oil or other foods with vitamin A, but little information is available on whether foods are actually adequately fortified in practice.3See this spreadsheet for the information we collected on the most recent vitamin A deficiency surveys and on vitamin A fortification programs in countries where HKI has supported vitamin A supplementation programs. jQuery("#footnote_plugin_tooltip_3").tooltip({ tip: "#footnote_plugin_tooltip_text_3", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

Taking the limited available data into account, our best guess is that prevalence of vitamin A deficiency in countries where HKI works today is likely to be considerably lower than the prevalence of vitamin A deficiency among children who participated in vitamin A supplementation trials—closer to 20% prevalence than 60% prevalence.

We find that HKI’s vitamin A supplementation programs still appear highly cost-effective, even when taking our estimate of the change in the prevalence of vitamin A deficiency over time into account (see our most recent cost-effectiveness analysis for full details). But we remain quite uncertain about our estimate of the prevalence of vitamin A deficiency in countries where HKI works—new information could cause us to update our views on HKI’s cost-effectiveness considerably.

Next year, we’ll continue to follow research relevant to estimating vitamin A deficiency rates where HKI works. We also plan to consider funding new vitamin A deficiency surveys ourselves through a GiveWell Incubation Grant.

Have improvements in health conditions over time reduced the need for vitamin A supplementation?

In a blog post last year, we wrote that vitamin A supplementation has a mixed evidence base. There is strong evidence from many randomized controlled trials conducted in the 1980s and 1990s that the program reduces child mortality, but a more recent trial in northern India with more participants than all the other trials combined (the Deworming and Enhanced Vitamin A trial, or DEVTA) did not find a statistically significant effect.

There have been broad declines in child mortality rates over the past few decades. Participants in the control group in the DEVTA trial had a mortality rate of 5.3 deaths per 1,000 child-years, lower than the mortality rates in the control groups in earlier trials that found statistically significant results (ranging from 10.6 to 126 deaths per 1,000 child-years). One potential explanation for the difference between the results of the DEVTA trial and earlier trials is that the some types of deaths prevented by vitamin A supplementation in previously studied populations had already been prevented through other means (e.g., increased access to immunizations and medical care) in the DEVTA population.

We looked into child mortality rates in countries in sub-Saharan Africa where HKI plans to use GiveWell-directed funding in the near future—Guinea, Burkina Faso, and Mali—as well as other countries where HKI has recently worked. Mortality rates among preschool-aged children in Guinea, Burkina Faso and Mali remain quite high—around 13 deaths per 1,000 child-years, within the range of mortality rates among control groups in vitamin A trials that found statistically significant results.4The control group mortality rate in the DEVTA trial was 5.3 per 1,000 child-years. See this spreadsheet for child mortality rates in Burkina Faso, Guinea, and Mali (13 deaths per 1,000 child-years is the simple average of “Average of GBD and UN IGME data” child mortality rates for the three countries), and see here for more information on control group mortality rates in other vitamin A supplementation trials. jQuery("#footnote_plugin_tooltip_4").tooltip({ tip: "#footnote_plugin_tooltip_text_4", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); Based on these high child mortality rates, we don’t believe it’s very likely that overall health conditions have improved enough in these countries for vitamin A supplementation to no longer be effective at preventing deaths.

It is also possible that changes in causes of child deaths between the 1980s and 1990s and today could mean that vitamin A supplementation is now less effective than it was in the past. Different vitamin A experts have different views on whether vitamin A primarily prevents deaths due to a few specific causes (we’ve seen diarrhea and measles most frequently pointed to) or whether it reduces deaths due to a wider range of conditions by, perhaps, strengthening the immune system against infection. In our view, the research on this is inconclusive. According to the data we’ve seen, infectious disease overall and diarrhea in particular cause a similar proportion of total deaths among young children today as they did in the 1980s and 1990s; measles causes a substantially lower proportion of total deaths today than it did in the past.5See the final bullet point in this section of our review of HKI for more on this topic. jQuery("#footnote_plugin_tooltip_5").tooltip({ tip: "#footnote_plugin_tooltip_text_5", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); We’ve added an adjustment to our cost-effectiveness analysis to account for changes in the composition of causes of child mortality since the vitamin A trials were implemented—HKI’s work still appears highly cost-effective following this adjustment.

We may conduct additional research next year to learn about child mortality rates in places where HKI works at a more granular (e.g., regional or sub-regional) level. We may also conduct additional research on the impact of changes in cause-specific mortality rates on the effectiveness of vitamin A supplementation.

How strong is HKI’s track record of supporting fixed-point vitamin A supplement distributions?

In many past HKI-supported campaigns, healthcare workers have traveled door-to-door to administer vitamin A supplements to preschool-aged children. Funding was already available from other sources for sending teams of healthcare workers door-to-door to administer polio vaccinations, and adding vitamin A supplementation to these campaigns was relatively simple and cheap.

In fixed-point distributions, caregivers are expected to bring their children to a central location to receive vitamin A supplements. Due to recent progress in polio elimination, many door-to-door programs have recently been scaled-down or eliminated, so HKI expects to primarily be supporting fixed-point distributions going forward.

It may be more challenging to reach a large proportion of a targeted population with fixed-point distributions. HKI’s recent monitoring surveys have found that, on average, its door-to-door distributions have achieved higher coverage rates (around 90%) than its fixed-point distributions (around 60%). The average of around 60% for fixed-point programs reflects surveys finding high coverage in a few campaigns in the Democratic Republic of the Congo and Mozambique, and relatively low coverage in campaigns in Nigeria, Tanzania, and Kenya.

A complication for assessing HKI’s track record is that HKI often chose to conduct coverage surveys in areas where it expected coverage to be particularly low, so we would guess that these results are not fully representative of HKI’s work on fixed-point distributions.

Based on the available information, our best guess is that HKI-supported fixed-point vitamin A supplementation distributions next year will achieve moderate to high coverage.6To be more precise about what I mean: in Guinea (the program I am most familiar with, following our site visit in October), I’m 70% confident that coverage surveys representative of the distribution as a whole will indicate that the first vitamin A supplement distribution in 2018 reached at least 55% of targeted children across the country. jQuery("#footnote_plugin_tooltip_6").tooltip({ tip: "#footnote_plugin_tooltip_text_6", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); HKI has told us that it will conduct representative monitoring surveys (not only in areas where it expects coverage to be low) following its vitamin A supplement distributions supported with GiveWell-directed funding next year—we expect that these surveys will provide data useful for assessing how successful the programs were overall.

Notes   [ + ]

1. ↑ See the “Imdad 2017 – VAD prevalence estimates” sheet here for details. 2. ↑ WHO Global prevalence of vitamin A deficiency in populations at risk 2009, Pg 8, Table 5. 3. ↑ See this spreadsheet for the information we collected on the most recent vitamin A deficiency surveys and on vitamin A fortification programs in countries where HKI has supported vitamin A supplementation programs. 4. ↑ The control group mortality rate in the DEVTA trial was 5.3 per 1,000 child-years. See this spreadsheet for child mortality rates in Burkina Faso, Guinea, and Mali (13 deaths per 1,000 child-years is the simple average of “Average of GBD and UN IGME data” child mortality rates for the three countries), and see here for more information on control group mortality rates in other vitamin A supplementation trials. 5. ↑ See the final bullet point in this section of our review of HKI for more on this topic. 6. ↑ To be more precise about what I mean: in Guinea (the program I am most familiar with, following our site visit in October), I’m 70% confident that coverage surveys representative of the distribution as a whole will indicate that the first vitamin A supplement distribution in 2018 reached at least 55% of targeted children across the country. function footnote_expand_reference_container() { jQuery("#footnote_references_container").show(); jQuery("#footnote_reference_container_collapse_button").text("-"); } function footnote_collapse_reference_container() { jQuery("#footnote_references_container").hide(); jQuery("#footnote_reference_container_collapse_button").text("+"); } function footnote_expand_collapse_reference_container() { if (jQuery("#footnote_references_container").is(":hidden")) { footnote_expand_reference_container(); } else { footnote_collapse_reference_container(); } } function footnote_moveToAnchor(p_str_TargetID) { footnote_expand_reference_container(); var l_obj_Target = jQuery("#" + p_str_TargetID); if(l_obj_Target.length) { jQuery('html, body').animate({ scrollTop: l_obj_Target.offset().top - window.innerHeight/2 }, 1000); } }

The post Key questions about Helen Keller International’s vitamin A supplementation program appeared first on The GiveWell Blog.

One of our two new top charities this year is Helen Keller International (HKI)’s vitamin A supplementation program. We named HKI’s vitamin A supplementation program a top charity this year because:

• There is strong evidence from many randomized controlled trials of vitamin A supplementation that the program leads to substantial reductions in child deaths.
• HKI-supported vitamin A supplementation programs are inexpensive (we estimate around $0.75 in total costs per supplement delivered) and highly cost-effective at preventing child deaths in countries where HKI plans to work using GiveWell-directed funds.
• HKI is transparent—it has shared significant, detailed information about its programs with us, including the results and methodology of monitoring surveys HKI conducted to determine whether its vitamin A supplementation programs reach a large proportion of targeted children.
• HKI has a funding gap—we believe it is highly likely that its vitamin A supplementation programs will be constrained by funding next year.

HKI’s vitamin A supplementation program is an exceptional giving opportunity, but as with the case for donating to any of our other top charities, not a “sure thing.”

I’m the Research Analyst who has led our work on HKI this year. In this post, I discuss some key questions about the impact of Helen Keller International’s vitamin A supplementation program and what we’ve learned so far. I also discuss GiveWell’s plans for learning more about these issues in the future.

In short:

• Is vitamin A deficiency still a major concern? Our best guess is that vitamin A deficiency is considerably less common today where HKI works than it was among children who participated in past trials of vitamin A supplementation, but not so rare that vitamin A supplementation would not be cost-effective. We are quite uncertain about our estimate of the prevalence of vitamin A deficiency where HKI works because little high-quality, up-to-date data on vitamin A deficiency is available. We plan to consider funding new surveys of vitamin A deficiency to improve our understanding of the effectiveness of HKI’s programs.
• Have improvements in health conditions over time reduced the need for vitamin A supplementation? Child mortality rates remain quite high in areas where HKI plans to use GiveWell-directed funding for vitamin A supplementation programs. We think it’s unlikely that health conditions in these countries have improved enough for vitamin A supplementation to no longer be effective.
• How strong is HKI’s track record of supporting fixed-point vitamin A supplement distributions? HKI expects to primarily support fixed-point vitamin A supplement distributions (rather than door-to-door campaigns) going forward. Results from monitoring surveys have found that, on average, HKI’s fixed-point programs have not reached as high a proportion of targeted populations as its door-to-door programs, but these monitoring surveys may not have been fully representative of HKI’s programs overall. Our best guess is that future fixed-point programs will achieve moderate to high coverage.

Is vitamin A deficiency still a major concern?

Vitamin A deficiency, a condition resulting from chronic low vitamin A intake, can cause loss of vision and increased severity of infections. If vitamin A deficiency is less common today than it was among participants in trials of vitamin A supplementation, today’s programs may prevent fewer deaths than the evidence from the trials suggests.

We estimate that the prevalence of vitamin A deficiency was high (around 60%) in the populations studied in trials included in the Cochrane Collaboration review of vitamin A supplementation programs for preschool-aged children, Imdad et al. 2017.1See the “Imdad 2017 – VAD prevalence estimates” sheet here for details. jQuery("#footnote_plugin_tooltip_1").tooltip({ tip: "#footnote_plugin_tooltip_text_1", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

The map below, from Our World in Data, presents the World Health Organization (WHO)’s most recent estimates of the prevalence of vitamin A deficiency among preschool-aged children by country, covering the period from 1995 to 2005. WHO categorizes prevalences of vitamin A deficiency among preschool-aged children of 20% or above as a severe public health problem.2WHO Global prevalence of vitamin A deficiency in populations at risk 2009, Pg 8, Table 5. jQuery("#footnote_plugin_tooltip_2").tooltip({ tip: "#footnote_plugin_tooltip_text_2", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

Since WHO’s most recent estimates are now considerably out-of-date, we decided to investigate a variety of additional sources in order to create best-guess estimates of rates of vitamin A deficiency today in countries in sub-Saharan Africa where HKI works.

We learned that there is very little useful, up-to-date data on vitamin A deficiency in countries in sub-Saharan Africa. In many countries, the most recent surveys of vitamin A deficiency were completed ten or more years ago. Many governments have also recently mandated the fortification of vegetable oil or other foods with vitamin A, but little information is available on whether foods are actually adequately fortified in practice.3See this spreadsheet for the information we collected on the most recent vitamin A deficiency surveys and on vitamin A fortification programs in countries where HKI has supported vitamin A supplementation programs. jQuery("#footnote_plugin_tooltip_3").tooltip({ tip: "#footnote_plugin_tooltip_text_3", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

Taking the limited available data into account, our best guess is that prevalence of vitamin A deficiency in countries where HKI works today is likely to be considerably lower than the prevalence of vitamin A deficiency among children who participated in vitamin A supplementation trials—closer to 20% prevalence than 60% prevalence.

We find that HKI’s vitamin A supplementation programs still appear highly cost-effective, even when taking our estimate of the change in the prevalence of vitamin A deficiency over time into account (see our most recent cost-effectiveness analysis for full details). But we remain quite uncertain about our estimate of the prevalence of vitamin A deficiency in countries where HKI works—new information could cause us to update our views on HKI’s cost-effectiveness considerably.

Next year, we’ll continue to follow research relevant to estimating vitamin A deficiency rates where HKI works. We also plan to consider funding new vitamin A deficiency surveys ourselves through a GiveWell Incubation Grant.

Have improvements in health conditions over time reduced the need for vitamin A supplementation?

In a blog post last year, we wrote that vitamin A supplementation has a mixed evidence base. There is strong evidence from many randomized controlled trials conducted in the 1980s and 1990s that the program reduces child mortality, but a more recent trial in northern India with more participants than all the other trials combined (the Deworming and Enhanced Vitamin A trial, or DEVTA) did not find a statistically significant effect.

There have been broad declines in child mortality rates over the past few decades. Participants in the control group in the DEVTA trial had a mortality rate of 5.3 deaths per 1,000 child-years, lower than the mortality rates in the control groups in earlier trials that found statistically significant results (ranging from 10.6 to 126 deaths per 1,000 child-years). One potential explanation for the difference between the results of the DEVTA trial and earlier trials is that the some types of deaths prevented by vitamin A supplementation in previously studied populations had already been prevented through other means (e.g., increased access to immunizations and medical care) in the DEVTA population.

We looked into child mortality rates in countries in sub-Saharan Africa where HKI plans to use GiveWell-directed funding in the near future—Guinea, Burkina Faso, and Mali—as well as other countries where HKI has recently worked. Mortality rates among preschool-aged children in Guinea, Burkina Faso and Mali remain quite high—around 13 deaths per 1,000 child-years, within the range of mortality rates among control groups in vitamin A trials that found statistically significant results.4The control group mortality rate in the DEVTA trial was 5.3 per 1,000 child-years. See this spreadsheet for child mortality rates in Burkina Faso, Guinea, and Mali (13 deaths per 1,000 child-years is the simple average of “Average of GBD and UN IGME data” child mortality rates for the three countries), and see here for more information on control group mortality rates in other vitamin A supplementation trials. jQuery("#footnote_plugin_tooltip_4").tooltip({ tip: "#footnote_plugin_tooltip_text_4", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); Based on these high child mortality rates, we don’t believe it’s very likely that overall health conditions have improved enough in these countries for vitamin A supplementation to no longer be effective at preventing deaths.

It is also possible that changes in causes of child deaths between the 1980s and 1990s and today could mean that vitamin A supplementation is now less effective than it was in the past. Different vitamin A experts have different views on whether vitamin A primarily prevents deaths due to a few specific causes (we’ve seen diarrhea and measles most frequently pointed to) or whether it reduces deaths due to a wider range of conditions by, perhaps, strengthening the immune system against infection. In our view, the research on this is inconclusive. According to the data we’ve seen, infectious disease overall and diarrhea in particular cause a similar proportion of total deaths among young children today as they did in the 1980s and 1990s; measles causes a substantially lower proportion of total deaths today than it did in the past.5See the final bullet point in this section of our review of HKI for more on this topic. jQuery("#footnote_plugin_tooltip_5").tooltip({ tip: "#footnote_plugin_tooltip_text_5", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); We’ve added an adjustment to our cost-effectiveness analysis to account for changes in the composition of causes of child mortality since the vitamin A trials were implemented—HKI’s work still appears highly cost-effective following this adjustment.

We may conduct additional research next year to learn about child mortality rates in places where HKI works at a more granular (e.g., regional or sub-regional) level. We may also conduct additional research on the impact of changes in cause-specific mortality rates on the effectiveness of vitamin A supplementation.

How strong is HKI’s track record of supporting fixed-point vitamin A supplement distributions?

In many past HKI-supported campaigns, healthcare workers have traveled door-to-door to administer vitamin A supplements to preschool-aged children. Funding was already available from other sources for sending teams of healthcare workers door-to-door to administer polio vaccinations, and adding vitamin A supplementation to these campaigns was relatively simple and cheap.

In fixed-point distributions, caregivers are expected to bring their children to a central location to receive vitamin A supplements. Due to recent progress in polio elimination, many door-to-door programs have recently been scaled-down or eliminated, so HKI expects to primarily be supporting fixed-point distributions going forward.

It may be more challenging to reach a large proportion of a targeted population with fixed-point distributions. HKI’s recent monitoring surveys have found that, on average, its door-to-door distributions have achieved higher coverage rates (around 90%) than its fixed-point distributions (around 60%). The average of around 60% for fixed-point programs reflects surveys finding high coverage in a few campaigns in the Democratic Republic of the Congo and Mozambique, and relatively low coverage in campaigns in Nigeria, Tanzania, and Kenya.

A complication for assessing HKI’s track record is that HKI often chose to conduct coverage surveys in areas where it expected coverage to be particularly low, so we would guess that these results are not fully representative of HKI’s work on fixed-point distributions.

Based on the available information, our best guess is that HKI-supported fixed-point vitamin A supplementation distributions next year will achieve moderate to high coverage.6To be more precise about what I mean: in Guinea (the program I am most familiar with, following our site visit in October), I’m 70% confident that coverage surveys representative of the distribution as a whole will indicate that the first vitamin A supplement distribution in 2018 reached at least 55% of targeted children across the country. jQuery("#footnote_plugin_tooltip_6").tooltip({ tip: "#footnote_plugin_tooltip_text_6", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); HKI has told us that it will conduct representative monitoring surveys (not only in areas where it expects coverage to be low) following its vitamin A supplement distributions supported with GiveWell-directed funding next year—we expect that these surveys will provide data useful for assessing how successful the programs were overall.

Notes   [ + ]

1. ↑ See the “Imdad 2017 – VAD prevalence estimates” sheet here for details. 2. ↑ WHO Global prevalence of vitamin A deficiency in populations at risk 2009, Pg 8, Table 5. 3. ↑ See this spreadsheet for the information we collected on the most recent vitamin A deficiency surveys and on vitamin A fortification programs in countries where HKI has supported vitamin A supplementation programs. 4. ↑ The control group mortality rate in the DEVTA trial was 5.3 per 1,000 child-years. See this spreadsheet for child mortality rates in Burkina Faso, Guinea, and Mali (13 deaths per 1,000 child-years is the simple average of “Average of GBD and UN IGME data” child mortality rates for the three countries), and see here for more information on control group mortality rates in other vitamin A supplementation trials. 5. ↑ See the final bullet point in this section of our review of HKI for more on this topic. 6. ↑ To be more precise about what I mean: in Guinea (the program I am most familiar with, following our site visit in October), I’m 70% confident that coverage surveys representative of the distribution as a whole will indicate that the first vitamin A supplement distribution in 2018 reached at least 55% of targeted children across the country. function footnote_expand_reference_container() { jQuery("#footnote_references_container").show(); jQuery("#footnote_reference_container_collapse_button").text("-"); } function footnote_collapse_reference_container() { jQuery("#footnote_references_container").hide(); jQuery("#footnote_reference_container_collapse_button").text("+"); } function footnote_expand_collapse_reference_container() { if (jQuery("#footnote_references_container").is(":hidden")) { footnote_expand_reference_container(); } else { footnote_collapse_reference_container(); } } function footnote_moveToAnchor(p_str_TargetID) { footnote_expand_reference_container(); var l_obj_Target = jQuery("#" + p_str_TargetID); if(l_obj_Target.length) { jQuery('html, body').animate({ scrollTop: l_obj_Target.offset().top - window.innerHeight/2 }, 1000); } }

The post Key questions about Helen Keller International’s vitamin A supplementation program appeared first on The GiveWell Blog.

When our cost-effectiveness analysis finds robust and meaningful differences between charities, it plays a large role in our recommendations (more on the role it plays in this post).

But while our cost-effectiveness analysis represent our best guess, it’s also subject to substantial uncertainty; some of its results are a function of highly debatable, difficult-to-estimate inputs.

Sometimes these inputs are largely subjective, such as the moral weight we assign to charities achieving different good outcomes (e.g. improving health vs increasing income). But even objective inputs are uncertain; a key input for anti-malaria interventions is malaria mortality, but the Institute for Health Metrics and Evaluation estimates 1.6 times more people died in Africa from malaria in 2016 (641,000) than the World Health Organization does (407,000; pg. 41).1Differences in their methodology have been discussed, with older figures, in a 2012 blog post by the Center for Global Development. jQuery("#footnote_plugin_tooltip_1").tooltip({ tip: "#footnote_plugin_tooltip_text_1", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

Before we finalized the recommendations we released in November, we determined how sensitive our results were to some of our most uncertain parameters.

In brief:

• Comparisons between charities achieving different types of good outcome are most sensitive to the relative value we assign to those outcomes (more on how and why we and other policymakers assign these weights in this post).
• Our deworming models are very uncertain, due to the complexity of the evidence base. They are also sensitive to the choice of discount rate: how we value good done today vs. good done in the future.
• Our malaria models (seasonal malaria chemoprevention and long-lasting insecticide-treated nets) are less uncertain than our deworming models, but are particularly sensitive to our estimate of the long term effects of malaria on income.

In this post, we discuss:

• The sensitivity of our analysis to moral weights (more) and other parameters (more).
• How this uncertainty influences our recommendations (more).
• Why this sensitivity analysis doesn’t capture the full scope of our uncertainty and ways in which we could improve our assessment and presentation of uncertainty (more).

The tornado charts at the bottom of this post show the results of our full sensitivity analysis. For a brief explanation of how we conducted our sensitivity analysis see this footnote.2Each contributor to our cost-effectiveness analysis inputs their own values for particularly uncertain parameters in our cost-effectiveness analysis. We use the median of contributors’ final cost-effectiveness results for our headline cost-effectiveness figures. To simplify the sensitivity analysis, we used the median of contributors’ parameter inputs to form a central cost-effectiveness estimate for each charity. The results below therefore differ slightly from our headline cost-effectiveness figures. To determine how sensitive the model is to each parameter, we flexed each parameter between the highest and lowest contributors’ inputs, while holding all other parameters constant. For more details, see our sensitivity analysis spreadsheet. jQuery("#footnote_plugin_tooltip_2").tooltip({ tip: "#footnote_plugin_tooltip_text_2", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

Sensitivity to moral weights

Some of the inputs in our model rely on judgement calls, which reasonable, informed people might disagree on. For example, we assign quantitative weights to our relative valuations of different good outcomes. These inputs capture important considerations in our decision-making, but are often difficult to justify precisely.

We ask contributors to our cost-effectiveness analysis (mostly staff) to input how many people’s income would have to double for 1 year to be equally valuable to averting the death of a child under 5. Contributors’ values vary widely, between 8 and 100 (see Figure 1).3You can see each of our contributors’ inputs for moral weights, and other uncertain parameters, on the Moral weights and Parameters tabs of our cost-effectiveness analysis. This year, contributors were also asked to provide a brief justification for their inputs in the cell notes. jQuery("#footnote_plugin_tooltip_3").tooltip({ tip: "#footnote_plugin_tooltip_text_3", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

Differences in cost-effectiveness between charities which primarily prevent child deaths (Helen Keller International, Malaria Consortium, Against Malaria Foundation) and charities which primarily increase income (Deworm the World Initiative, Schistosomiasis Control Initiative, Sightsavers, No Lean Season, End Fund) are highly sensitive to different plausible moral weights (See Figure 2).

The orange points represent the median estimated cost-effectiveness of our charities (in terms of how many times more cost-effective than GiveDirectly we model them to be). The blue bars represents the range of cost-effectiveness for different valuations of preventing the death of an under-5 child between 8x and 100x as good as doubling consumption for one person for one year (holding all other parameters in the model constant). Deworming sensitivities

Our deworming models are very uncertain, due to the complexity of the evidence base, and the long time horizons over which we expect the potential benefits to be realized. Aside from our moral weights, our deworming charities are highly sensitive to three inputs:

• Replicability adjustment. We make a “replicability adjustment” for deworming to account for the fact that the consumption increase in a major study we rely upon may not hold up if it were replicated. If you’re skeptical that such a large income increase would occur, given the limited evidence for short-term health benefits and generally unexpected nature of the findings, you may think that the effect the study measured wasn’t real, wasn’t driven by deworming, or relied on an atypical characteristic shared by the study population but not likely to found among recipients of the intervention today. This adjustment is not well-grounded in data. (For more discussion see our deworming intervention report and blog posts here, here, here and here).4You can read more about how contributors settled on the values they used for this parameter in the cell notes in row 16 of the Parameters sheet of our November 2017 cost-effectiveness model. jQuery("#footnote_plugin_tooltip_4").tooltip({ tip: "#footnote_plugin_tooltip_text_4", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });
• Adjustment for years of treatment in Baird et al. vs. years of treatment in charities’ programs. Our charities aim to deworm children for up to 10 years, which is longer than the intervention studies in Baird et al. 2015 (where children in the treatment group received 4 years of deworming). There may be diminishing returns as the total years of treatment increase, although this is difficult to estimate.
• Discount rate. The discount rate adjusts for benefits that occur at different points in time. For a number of reasons, individuals may believe it is preferable for income to rise now than at some point in the future.

Figure 3 shows how the cost-effectiveness of Deworm the World Initiative5The sensitivity of other deworming charities is largely dependent on the same parameters. Charts are presented in the Appendix jQuery("#footnote_plugin_tooltip_5").tooltip({ tip: "#footnote_plugin_tooltip_text_5", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); varies depending on different contributor inputs for different parameters (more on how to interpret these parameters here).

The orange line represents the median estimated cost-effectiveness of our charities (in terms of how many times more cost-effective than GiveDirectly we model them to be). The blue bars represents the range of cost-effectiveness for different inputs from our contributors for that parameter (holding all other parameters in the model constant). The figures in square brackets represent the range of contributor inputs for those parameters. Malaria sensitivities

Our malaria models are less uncertain than our deworming models, but are still sensitive to our estimate of the long term effects of malaria on income (see Figures 4 and 5).

Interpreting the evidence base for the effect of malaria prevention on long run income is complex, and contributors differ widely in their interpretation. We’re planning to do more research on this topic further but summarize our current understanding here.

What does this mean for our recommendations?

When we model large differences in cost-effectiveness, we generally follow those recommendations. When charities are closer on cost-effectiveness, we pay more attention to qualitative considerations, such as the quality of their monitoring and evaluation, and potential upside benefits which are difficult to quantify (e.g. scaling a novel program).

What counts as a meaningful difference in modelled cost-effectiveness depends on a number of factors, including:

• Do the programs target the same outcomes? We place less weight on modelled differences between charities which have different good outcomes because our cost-effectiveness analysis is sensitive to different reasonable moral weights.
• How similar are the programs? We’re more confident in our comparison between our deworming charities than we are between deworming charities and other charities targeting income such as GiveDirectly. This is because we expect the most likely errors in our deworming estimates (e.g. based on our interpretation of the evidence) for different charities to be correlated.
• Are there important qualitative reasons to differentiate between the charities? We place less relative weight on cost-effectiveness analysis when there are important qualitative reasons to differentiate between charities.

For a more detailed explanation of how we made our recommendations this year, see our recent announcement of our top charities for giving season 2017.

What are the limitations of this sensitivity analysis?

This sensitivity analysis shouldn’t be taken as a full representation of our all things considered uncertainty:

• The charts above show the sensitivity of the cost-effectiveness analysis to changing one input at a time (holding all other constant). The ranges don’t necessarily imply any particular credible interval, and are more useful for identifying which inputs are most uncertain than for reflecting our all things considered uncertainty around the cost-effectiveness of a particular charity.
• We don’t ask multiple contributors to input their own values for all uncertain inputs (e.g. because we think the benefits of using the inputs of the contributors with most context outweigh the benefit of getting inputs from many contributors). These inputs have not been included in the sensitivity analysis.
• Model uncertainty. Explicitly modelling all the considerations relevant to our charity would be infeasible. Even if all our inputs were fully accurate, we’d still retain some uncertainty about the true cost-effectiveness of our charities.

We’re considering a number of different options to improve our sensitivity analysis and communication of uncertainty in the future, such as expressing inputs as probability distributions or creating a Monte Carlo simulation. But we’re uncertain whether these would create sufficient decision-relevant information for our readers to justify the substantial time investment and additional complexity.

If you’d find such an analysis helpful, let us know in the comments.

Appendix

In this section, we present tornado charts for each of our top charities. You can see more detailed descriptions of how to interpret these parameters here, or in the cell notes of our cost-effectiveness analysis.

Notes   [ + ]

1. ↑ Differences in their methodology have been discussed, with older figures, in a 2012 blog post by the Center for Global Development. 2. ↑ Each contributor to our cost-effectiveness analysis inputs their own values for particularly uncertain parameters in our cost-effectiveness analysis. We use the median of contributors’ final cost-effectiveness results for our headline cost-effectiveness figures. To simplify the sensitivity analysis, we used the median of contributors’ parameter inputs to form a central cost-effectiveness estimate for each charity. The results below therefore differ slightly from our headline cost-effectiveness figures. To determine how sensitive the model is to each parameter, we flexed each parameter between the highest and lowest contributors’ inputs, while holding all other parameters constant. For more details, see our sensitivity analysis spreadsheet. 3. ↑ You can see each of our contributors’ inputs for moral weights, and other uncertain parameters, on the Moral weights and Parameters tabs of our cost-effectiveness analysis. This year, contributors were also asked to provide a brief justification for their inputs in the cell notes. 4. ↑ You can read more about how contributors settled on the values they used for this parameter in the cell notes in row 16 of the Parameters sheet of our November 2017 cost-effectiveness model. 5. ↑ The sensitivity of other deworming charities is largely dependent on the same parameters. Charts are presented in the Appendix function footnote_expand_reference_container() { jQuery("#footnote_references_container").show(); jQuery("#footnote_reference_container_collapse_button").text("-"); } function footnote_collapse_reference_container() { jQuery("#footnote_references_container").hide(); jQuery("#footnote_reference_container_collapse_button").text("+"); } function footnote_expand_collapse_reference_container() { if (jQuery("#footnote_references_container").is(":hidden")) { footnote_expand_reference_container(); } else { footnote_collapse_reference_container(); } } function footnote_moveToAnchor(p_str_TargetID) { footnote_expand_reference_container(); var l_obj_Target = jQuery("#" + p_str_TargetID); if(l_obj_Target.length) { jQuery('html, body').animate({ scrollTop: l_obj_Target.offset().top - window.innerHeight/2 }, 1000); } }

The post How uncertain is our cost-effectiveness analysis? appeared first on The GiveWell Blog.

When our cost-effectiveness analysis finds robust and meaningful differences between charities, it plays a large role in our recommendations (more on the role it plays in this post).

But while our cost-effectiveness analysis represent our best guess, it’s also subject to substantial uncertainty; some of its results are a function of highly debatable, difficult-to-estimate inputs.

Sometimes these inputs are largely subjective, such as the moral weight we assign to charities achieving different good outcomes (e.g. improving health vs increasing income). But even objective inputs are uncertain; a key input for anti-malaria interventions is malaria mortality, but the Institute for Health Metrics and Evaluation estimates 1.6 times more people died in Africa from malaria in 2016 (641,000) than the World Health Organization does (407,000; pg. 41).1Differences in their methodology have been discussed, with older figures, in a 2012 blog post by the Center for Global Development. jQuery("#footnote_plugin_tooltip_1").tooltip({ tip: "#footnote_plugin_tooltip_text_1", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

Before we finalized the recommendations we released in November, we determined how sensitive our results were to some of our most uncertain parameters.

In brief:

• Comparisons between charities achieving different types of good outcome are most sensitive to the relative value we assign to those outcomes (more on how and why we and other policymakers assign these weights in this post).
• Our deworming models are very uncertain, due to the complexity of the evidence base. They are also sensitive to the choice of discount rate: how we value good done today vs. good done in the future.
• Our malaria models (seasonal malaria chemoprevention and long-lasting insecticide-treated nets) are less uncertain than our deworming models, but are particularly sensitive to our estimate of the long term effects of malaria on income.

In this post, we discuss:

• The sensitivity of our analysis to moral weights (more) and other parameters (more).
• How this uncertainty influences our recommendations (more).
• Why this sensitivity analysis doesn’t capture the full scope of our uncertainty and ways in which we could improve our assessment and presentation of uncertainty (more).

The tornado charts at the bottom of this post show the results of our full sensitivity analysis. For a brief explanation of how we conducted our sensitivity analysis see this footnote.2Each contributor to our cost-effectiveness analysis inputs their own values for particularly uncertain parameters in our cost-effectiveness analysis. We use the median of contributors’ final cost-effectiveness results for our headline cost-effectiveness figures. To simplify the sensitivity analysis, we used the median of contributors’ parameter inputs to form a central cost-effectiveness estimate for each charity. The results below therefore differ slightly from our headline cost-effectiveness figures. To determine how sensitive the model is to each parameter, we flexed each parameter between the highest and lowest contributors’ inputs, while holding all other parameters constant. For more details, see our sensitivity analysis spreadsheet. jQuery("#footnote_plugin_tooltip_2").tooltip({ tip: "#footnote_plugin_tooltip_text_2", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

Sensitivity to moral weights

Some of the inputs in our model rely on judgement calls, which reasonable, informed people might disagree on. For example, we assign quantitative weights to our relative valuations of different good outcomes. These inputs capture important considerations in our decision-making, but are often difficult to justify precisely.

We ask contributors to our cost-effectiveness analysis (mostly staff) to input how many people’s income would have to double for 1 year to be equally valuable to averting the death of a child under 5. Contributors’ values vary widely, between 8 and 100 (see Figure 1).3You can see each of our contributors’ inputs for moral weights, and other uncertain parameters, on the Moral weights and Parameters tabs of our cost-effectiveness analysis. This year, contributors were also asked to provide a brief justification for their inputs in the cell notes. jQuery("#footnote_plugin_tooltip_3").tooltip({ tip: "#footnote_plugin_tooltip_text_3", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

Differences in cost-effectiveness between charities which primarily prevent child deaths (Helen Keller International, Malaria Consortium, Against Malaria Foundation) and charities which primarily increase income (Deworm the World Initiative, Schistosomiasis Control Initiative, Sightsavers, No Lean Season, End Fund) are highly sensitive to different plausible moral weights (See Figure 2).

The orange points represent the median estimated cost-effectiveness of our charities (in terms of how many times more cost-effective than GiveDirectly we model them to be). The blue bars represents the range of cost-effectiveness for different valuations of preventing the death of an under-5 child between 8x and 100x as good as doubling consumption for one person for one year (holding all other parameters in the model constant). Deworming sensitivities

Our deworming models are very uncertain, due to the complexity of the evidence base, and the long time horizons over which we expect the potential benefits to be realized. Aside from our moral weights, our deworming charities are highly sensitive to three inputs:

• Replicability adjustment. We make a “replicability adjustment” for deworming to account for the fact that the consumption increase in a major study we rely upon may not hold up if it were replicated. If you’re skeptical that such a large income increase would occur, given the limited evidence for short-term health benefits and generally unexpected nature of the findings, you may think that the effect the study measured wasn’t real, wasn’t driven by deworming, or relied on an atypical characteristic shared by the study population but not likely to found among recipients of the intervention today. This adjustment is not well-grounded in data. (For more discussion see our deworming intervention report and blog posts here, here, here and here).4You can read more about how contributors settled on the values they used for this parameter in the cell notes in row 16 of the Parameters sheet of our November 2017 cost-effectiveness model. jQuery("#footnote_plugin_tooltip_4").tooltip({ tip: "#footnote_plugin_tooltip_text_4", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });
• Adjustment for years of treatment in Baird et al. vs. years of treatment in charities’ programs. Our charities aim to deworm children for up to 10 years, which is longer than the intervention studies in Baird et al. 2015 (where children in the treatment group received 4 years of deworming). There may be diminishing returns as the total years of treatment increase, although this is difficult to estimate.
• Discount rate. The discount rate adjusts for benefits that occur at different points in time. For a number of reasons, individuals may believe it is preferable for income to rise now than at some point in the future.

Figure 3 shows how the cost-effectiveness of Deworm the World Initiative5The sensitivity of other deworming charities is largely dependent on the same parameters. Charts are presented in the Appendix jQuery("#footnote_plugin_tooltip_5").tooltip({ tip: "#footnote_plugin_tooltip_text_5", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); varies depending on different contributor inputs for different parameters (more on how to interpret these parameters here).

The orange line represents the median estimated cost-effectiveness of our charities (in terms of how many times more cost-effective than GiveDirectly we model them to be). The blue bars represents the range of cost-effectiveness for different inputs from our contributors for that parameter (holding all other parameters in the model constant). The figures in square brackets represent the range of contributor inputs for those parameters. Malaria sensitivities

Our malaria models are less uncertain than our deworming models, but are still sensitive to our estimate of the long term effects of malaria on income (see Figures 4 and 5).

Interpreting the evidence base for the effect of malaria prevention on long run income is complex, and contributors differ widely in their interpretation. We’re planning to do more research on this topic further but summarize our current understanding here.

What does this mean for our recommendations?

When we model large differences in cost-effectiveness, we generally follow those recommendations. When charities are closer on cost-effectiveness, we pay more attention to qualitative considerations, such as the quality of their monitoring and evaluation, and potential upside benefits which are difficult to quantify (e.g. scaling a novel program).

What counts as a meaningful difference in modelled cost-effectiveness depends on a number of factors, including:

• Do the programs target the same outcomes? We place less weight on modelled differences between charities which have different good outcomes because our cost-effectiveness analysis is sensitive to different reasonable moral weights.
• How similar are the programs? We’re more confident in our comparison between our deworming charities than we are between deworming charities and other charities targeting income such as GiveDirectly. This is because we expect the most likely errors in our deworming estimates (e.g. based on our interpretation of the evidence) for different charities to be correlated.
• Are there important qualitative reasons to differentiate between the charities? We place less relative weight on cost-effectiveness analysis when there are important qualitative reasons to differentiate between charities.

For a more detailed explanation of how we made our recommendations this year, see our recent announcement of our top charities for giving season 2017.

What are the limitations of this sensitivity analysis?

This sensitivity analysis shouldn’t be taken as a full representation of our all things considered uncertainty:

• The charts above show the sensitivity of the cost-effectiveness analysis to changing one input at a time (holding all other constant). The ranges don’t necessarily imply any particular credible interval, and are more useful for identifying which inputs are most uncertain than for reflecting our all things considered uncertainty around the cost-effectiveness of a particular charity.
• We don’t ask multiple contributors to input their own values for all uncertain inputs (e.g. because we think the benefits of using the inputs of the contributors with most context outweigh the benefit of getting inputs from many contributors). These inputs have not been included in the sensitivity analysis.
• Model uncertainty. Explicitly modelling all the considerations relevant to our charity would be infeasible. Even if all our inputs were fully accurate, we’d still retain some uncertainty about the true cost-effectiveness of our charities.

We’re considering a number of different options to improve our sensitivity analysis and communication of uncertainty in the future, such as expressing inputs as probability distributions or creating a Monte Carlo simulation. But we’re uncertain whether these would create sufficient decision-relevant information for our readers to justify the substantial time investment and additional complexity.

If you’d find such an analysis helpful, let us know in the comments.

Appendix

In this section, we present tornado charts for each of our top charities. You can see more detailed descriptions of how to interpret these parameters here, or in the cell notes of our cost-effectiveness analysis.

Notes   [ + ]

1. ↑ Differences in their methodology have been discussed, with older figures, in a 2012 blog post by the Center for Global Development. 2. ↑ Each contributor to our cost-effectiveness analysis inputs their own values for particularly uncertain parameters in our cost-effectiveness analysis. We use the median of contributors’ final cost-effectiveness results for our headline cost-effectiveness figures. To simplify the sensitivity analysis, we used the median of contributors’ parameter inputs to form a central cost-effectiveness estimate for each charity. The results below therefore differ slightly from our headline cost-effectiveness figures. To determine how sensitive the model is to each parameter, we flexed each parameter between the highest and lowest contributors’ inputs, while holding all other parameters constant. For more details, see our sensitivity analysis spreadsheet. 3. ↑ You can see each of our contributors’ inputs for moral weights, and other uncertain parameters, on the Moral weights and Parameters tabs of our cost-effectiveness analysis. This year, contributors were also asked to provide a brief justification for their inputs in the cell notes. 4. ↑ You can read more about how contributors settled on the values they used for this parameter in the cell notes in row 16 of the Parameters sheet of our November 2017 cost-effectiveness model. 5. ↑ The sensitivity of other deworming charities is largely dependent on the same parameters. Charts are presented in the Appendix function footnote_expand_reference_container() { jQuery("#footnote_references_container").show(); jQuery("#footnote_reference_container_collapse_button").text("-"); } function footnote_collapse_reference_container() { jQuery("#footnote_references_container").hide(); jQuery("#footnote_reference_container_collapse_button").text("+"); } function footnote_expand_collapse_reference_container() { if (jQuery("#footnote_references_container").is(":hidden")) { footnote_expand_reference_container(); } else { footnote_collapse_reference_container(); } } function footnote_moveToAnchor(p_str_TargetID) { footnote_expand_reference_container(); var l_obj_Target = jQuery("#" + p_str_TargetID); if(l_obj_Target.length) { jQuery('html, body').animate({ scrollTop: l_obj_Target.offset().top - window.innerHeight/2 }, 1000); } }

The post How uncertain is our cost-effectiveness analysis? appeared first on The GiveWell Blog.

GiveWell’s impact is a function of the quality of our research and the amount of money we direct to our recommended charities (our “money moved”). Historically, we’ve focused mostly on research because we felt that the quality of our recommendations was a greater constraint to our impact than our money moved.

This has changed. Outreach is now a major organizational priority. The goal of this work is to increase the amount of money we direct to our top-recommended charities.

In April 2014 I wrote about our work on outreach to explain why we hadn’t prioritized it: in brief, our growth had largely been driven by inbound interest in GiveWell, and proactive outreach efforts (beyond building relationships with existing donors) hadn’t yielded results that were worth the cost.

What changed?

• We believe that the amount of money we move is now a greater constraint to our impact than additional improvements in the quality of our research. Over the last two years, we’ve added five new top charities (three of which implement programs that weren’t previously represented on our top charities list), and we expect that our top charities, collectively, will have more than $200 million in unfilled funding gaps once they’ve received the funding that we expect to direct to them. (This calculation excludes GiveDirectly, which we believe could absorb and distribute hundreds of millions of dollars.) At the same time, the quality of our research and our capacity for research is higher than it’s ever been, so the returns to adding staff there (in terms of the pace at which we identify significantly better giving opportunities) are now lower.
• Increased capacity for outreach. In our 2014 post, we wrote that one of our key constraints was that senior staff (which at the time meant primarily GiveWell Co-Founder Holden Karnofsky and me) were necessary for most outreach-related work. This has changed. We now have capacity to take on outreach work as other staff have been hired and trained on this type of work.
• Better information on the impact of GiveWell’s outreach. We now have better information about the returns to outreach because:
  1. We’ve collected better data (via an improved donations processing system and outreach efforts) about where donors find out about us. Because of our ability to track donors, we know that a single appearance on NPR or major podcasts tends to drive $50,000+ in annual donations.
  2. More time passing has demonstrated that the lifetime value of the donations of a first time donor is higher than we expected. In several cases, we’ve seen major donors (i.e., those giving $10,000-$100,000) increase their annual giving by a factor of 10 or more.

We’re in the early stages of figuring out how we can proactively invest time and money in outreach to significantly increase our money moved. For now, we’ve taken some opportunities that we think will have positive returns; these are the three that we’ve invested the most time and money in to date:

• Podcast advertising. We’ve been advertising on podcasts that we believe our target audience listens to, based on interviews with current donors and GiveWell staff. In February and March, we ran a small experiment with a few ads on FiveThirtyEight’s Politics podcast and Vox’s The Weeds.1We’ve also been running ads on Julia Galef’s Rationally Speaking podcast since then. Because it’s much smaller and more targeted, we’ve excluded it from this analysis. Measured returns to advertising on Rationally Speaking have been significantly better than the more mainstream podcasts discussed in this post. jQuery("#footnote_plugin_tooltip_1").tooltip({ tip: "#footnote_plugin_tooltip_text_1", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

  In total, we spent approximately $20,000 on ads for this initial experiment. We ask donors who give via our website to tell us where they learned about GiveWell when they donate. GiveWell received approximately $8,000 in donations between February 1 and November 20 from donors who reported that they had learned about us via these podcasts.

  The donations we received were from first-time donors; to assess the impact of our advertising, we need to estimate the lifetime value of acquiring a new donor. In work we’ve done to assess our retention rate, we’ve seen that (a) approximately 20-25% of the donors who make a first-time donation of less than $1,000 give again in the subsequent year but (b) because many first-time donors increase the size of their donation over time, collectively, the donors who recur give more than 100% of the value of what they give in their first year.

  At higher donation levels ($1,000-$100,000), we measure 40-45% retention among donors, which leads to retention of approximately two-thirds of dollars given.2I say “measure” retention because we’ve learned that many donors give subsequent donations directly to our top charities and don’t report those donations to us. We’ve tried to follow up with lapsed donors and with charities to track these donors down. jQuery("#footnote_plugin_tooltip_2").tooltip({ tip: "#footnote_plugin_tooltip_text_2", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

  We therefore estimated the net present value of expected future donations (over the next five years) from these podcasts ads as somewhere between approximately $20,000 (assuming two-thirds dollar retention for the first two years and 100% dollar retention subsequently) and $45,000 (assuming 100% dollar retention).3We only projected donations over five years. This is fairly arbitrary because we don’t have long-term enough data to know whether or not this is a reasonable assumption. We capped it to prevent our assessment being driven by speculation about how much money would be donated many years in the future. jQuery("#footnote_plugin_tooltip_3").tooltip({ tip: "#footnote_plugin_tooltip_text_3", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

  A few additional facts are worth keeping in mind about the above figures:

  • We ran this experiment in February and March; most donors give at the end of the calendar year. We consistently see donors who find out about GiveWell during the course of the year, but donate in December. Other things equal, we expect that our advertising would have had greater measured returns in December than earlier in the year.
  • We are only able to track donors who (a) fill out our donation form telling us where they learned about us and (b) give directly through our website rather than to our top charities. Less than 50% of donors who give via credit card (and a smaller percentage of donors who give via check) tell us where they learned about GiveWell. Also, roughly speaking, approximately 50% of the donors and dollars we influence come through GiveWell rather than going to our top charities.4I took this rough estimate from footnote 26, on page 15, of GiveWell’s 2015 metrics report. jQuery("#footnote_plugin_tooltip_4").tooltip({ tip: "#footnote_plugin_tooltip_text_4", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });
  • It’s certainly possible that donors who learn about us via podcast would be more likely to give through our website than an average donor, more likely to report on how they found us (since their source is clear), or less likely to be retained. My best guess is that donors who learn about us via podcast ads behave similarly to our other donors, but I won’t be surprised if they don’t.

  With all that in mind, I believe that the impact of our podcast advertising is higher than what we directly measured.

  The results we saw from February to November this year were promising enough that we decided to increase the size of our experiment by spending approximately $100,000 on podcast ads. We’re currently running ads on FiveThirtyEight’s Politics podcast and Ezra Klein’s podcast and The Weeds at Vox.

• Earned media outreach. Mentions of GiveWell in the media have historically been a strong driver of growth. We aimed to increase mentions of GiveWell in high-quality, high-profile media where we’ve had the most past success as measured by dollars donated (i.e., media like The New York Times, NPR, The Wall Street Journal, and Financial Times). We retained a PR firm that came strongly recommended; we also increased 1-to-1 outreach by GiveWell staff to members of the media who have covered GiveWell in the past. It’s very hard to attribute the impact of the additional effort we’ve invested—overall, our effort has been fairly limited, and it’s hard to easily draw the causal lines between our work and the stories that appear—but my guess is that our increased efforts have led to more coverage of GiveWell and our top charities this giving season than in the recent past.
• Website improvements. Companies that sell products online invest significant effort into optimizing their websites and checkout pages to maximize their revenues. We retained a marketing consultant, Will Wong of Mission Street, and we’ve been A/B testing different donation pages and plan to test other pages on our website such as our homepage or top charities page to see whether we can increase our conversion rate (i.e., the percentage of visitors to our website who give to one of our top charities). For context, our current conversion rate is 1%. Our understanding is that a standard conversion rate for e-commerce companies is 2%, and that international nonprofits have a similar conversion rate.5See Pg 51 of the study downloadable here. jQuery("#footnote_plugin_tooltip_5").tooltip({ tip: "#footnote_plugin_tooltip_text_5", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); An increase in our conversion rate to the industry average would lead to a significant increase in the amount of money we direct to our top charities.

Notes   [ + ]

1. ↑ We’ve also been running ads on Julia Galef’s Rationally Speaking podcast since then. Because it’s much smaller and more targeted, we’ve excluded it from this analysis. Measured returns to advertising on Rationally Speaking have been significantly better than the more mainstream podcasts discussed in this post. 2. ↑ I say “measure” retention because we’ve learned that many donors give subsequent donations directly to our top charities and don’t report those donations to us. We’ve tried to follow up with lapsed donors and with charities to track these donors down. 3. ↑ We only projected donations over five years. This is fairly arbitrary because we don’t have long-term enough data to know whether or not this is a reasonable assumption. We capped it to prevent our assessment being driven by speculation about how much money would be donated many years in the future. 4. ↑ I took this rough estimate from footnote 26, on page 15, of GiveWell’s 2015 metrics report. 5. ↑ See Pg 51 of the study downloadable here. function footnote_expand_reference_container() { jQuery("#footnote_references_container").show(); jQuery("#footnote_reference_container_collapse_button").text("-"); } function footnote_collapse_reference_container() { jQuery("#footnote_references_container").hide(); jQuery("#footnote_reference_container_collapse_button").text("+"); } function footnote_expand_collapse_reference_container() { if (jQuery("#footnote_references_container").is(":hidden")) { footnote_expand_reference_container(); } else { footnote_collapse_reference_container(); } } function footnote_moveToAnchor(p_str_TargetID) { footnote_expand_reference_container(); var l_obj_Target = jQuery("#" + p_str_TargetID); if(l_obj_Target.length) { jQuery('html, body').animate({ scrollTop: l_obj_Target.offset().top - window.innerHeight/2 }, 1000); } }

The post Update on our work on outreach appeared first on The GiveWell Blog.

GiveWell’s impact is a function of the quality of our research and the amount of money we direct to our recommended charities (our “money moved”). Historically, we’ve focused mostly on research because we felt that the quality of our recommendations was a greater constraint to our impact than our money moved.

This has changed. Outreach is now a major organizational priority. The goal of this work is to increase the amount of money we direct to our top-recommended charities.

In April 2014 I wrote about our work on outreach to explain why we hadn’t prioritized it: in brief, our growth had largely been driven by inbound interest in GiveWell, and proactive outreach efforts (beyond building relationships with existing donors) hadn’t yielded results that were worth the cost.

What changed?

• We believe that the amount of money we move is now a greater constraint to our impact than additional improvements in the quality of our research. Over the last two years, we’ve added five new top charities (three of which implement programs that weren’t previously represented on our top charities list), and we expect that our top charities, collectively, will have more than $200 million in unfilled funding gaps once they’ve received the funding that we expect to direct to them. (This calculation excludes GiveDirectly, which we believe could absorb and distribute hundreds of millions of dollars.) At the same time, the quality of our research and our capacity for research is higher than it’s ever been, so the returns to adding staff there (in terms of the pace at which we identify significantly better giving opportunities) are now lower.
• Increased capacity for outreach. In our 2014 post, we wrote that one of our key constraints was that senior staff (which at the time meant primarily GiveWell Co-Founder Holden Karnofsky and me) were necessary for most outreach-related work. This has changed. We now have capacity to take on outreach work as other staff have been hired and trained on this type of work.
• Better information on the impact of GiveWell’s outreach. We now have better information about the returns to outreach because:
  1. We’ve collected better data (via an improved donations processing system and outreach efforts) about where donors find out about us. Because of our ability to track donors, we know that a single appearance on NPR or major podcasts tends to drive $50,000+ in annual donations.
  2. More time passing has demonstrated that the lifetime value of the donations of a first time donor is higher than we expected. In several cases, we’ve seen major donors (i.e., those giving $10,000-$100,000) increase their annual giving by a factor of 10 or more.

We’re in the early stages of figuring out how we can proactively invest time and money in outreach to significantly increase our money moved. For now, we’ve taken some opportunities that we think will have positive returns; these are the three that we’ve invested the most time and money in to date:

• Podcast advertising. We’ve been advertising on podcasts that we believe our target audience listens to, based on interviews with current donors and GiveWell staff. In February and March, we ran a small experiment with a few ads on FiveThirtyEight’s Politics podcast and Vox’s The Weeds.1We’ve also been running ads on Julia Galef’s Rationally Speaking podcast since then. Because it’s much smaller and more targeted, we’ve excluded it from this analysis. Measured returns to advertising on Rationally Speaking have been significantly better than the more mainstream podcasts discussed in this post. jQuery("#footnote_plugin_tooltip_1").tooltip({ tip: "#footnote_plugin_tooltip_text_1", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

  In total, we spent approximately $20,000 on ads for this initial experiment. We ask donors who give via our website to tell us where they learned about GiveWell when they donate. GiveWell received approximately $8,000 in donations between February 1 and November 20 from donors who reported that they had learned about us via these podcasts.

  The donations we received were from first-time donors; to assess the impact of our advertising, we need to estimate the lifetime value of acquiring a new donor. In work we’ve done to assess our retention rate, we’ve seen that (a) approximately 20-25% of the donors who make a first-time donation of less than $1,000 give again in the subsequent year but (b) because many first-time donors increase the size of their donation over time, collectively, the donors who recur give more than 100% of the value of what they give in their first year.

  At higher donation levels ($1,000-$100,000), we measure 40-45% retention among donors, which leads to retention of approximately two-thirds of dollars given.2I say “measure” retention because we’ve learned that many donors give subsequent donations directly to our top charities and don’t report those donations to us. We’ve tried to follow up with lapsed donors and with charities to track these donors down. jQuery("#footnote_plugin_tooltip_2").tooltip({ tip: "#footnote_plugin_tooltip_text_2", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

  We therefore estimated the net present value of expected future donations (over the next five years) from these podcasts ads as somewhere between approximately $20,000 (assuming two-thirds dollar retention for the first two years and 100% dollar retention subsequently) and $45,000 (assuming 100% dollar retention).3We only projected donations over five years. This is fairly arbitrary because we don’t have long-term enough data to know whether or not this is a reasonable assumption. We capped it to prevent our assessment being driven by speculation about how much money would be donated many years in the future. jQuery("#footnote_plugin_tooltip_3").tooltip({ tip: "#footnote_plugin_tooltip_text_3", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

  A few additional facts are worth keeping in mind about the above figures:

  • We ran this experiment in February and March; most donors give at the end of the calendar year. We consistently see donors who find out about GiveWell during the course of the year, but donate in December. Other things equal, we expect that our advertising would have had greater measured returns in December than earlier in the year.
  • We are only able to track donors who (a) fill out our donation form telling us where they learned about us and (b) give directly through our website rather than to our top charities. Less than 50% of donors who give via credit card (and a smaller percentage of donors who give via check) tell us where they learned about GiveWell. Also, roughly speaking, approximately 50% of the donors and dollars we influence come through GiveWell rather than going to our top charities.4I took this rough estimate from footnote 26, on page 15, of GiveWell’s 2015 metrics report. jQuery("#footnote_plugin_tooltip_4").tooltip({ tip: "#footnote_plugin_tooltip_text_4", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });
  • It’s certainly possible that donors who learn about us via podcast would be more likely to give through our website than an average donor, more likely to report on how they found us (since their source is clear), or less likely to be retained. My best guess is that donors who learn about us via podcast ads behave similarly to our other donors, but I won’t be surprised if they don’t.

  With all that in mind, I believe that the impact of our podcast advertising is higher than what we directly measured.

  The results we saw from February to November this year were promising enough that we decided to increase the size of our experiment by spending approximately $100,000 on podcast ads. We’re currently running ads on FiveThirtyEight’s Politics podcast and Ezra Klein’s podcast and The Weeds at Vox.

• Earned media outreach. Mentions of GiveWell in the media have historically been a strong driver of growth. We aimed to increase mentions of GiveWell in high-quality, high-profile media where we’ve had the most past success as measured by dollars donated (i.e., media like The New York Times, NPR, The Wall Street Journal, and Financial Times). We retained a PR firm that came strongly recommended; we also increased 1-to-1 outreach by GiveWell staff to members of the media who have covered GiveWell in the past. It’s very hard to attribute the impact of the additional effort we’ve invested—overall, our effort has been fairly limited, and it’s hard to easily draw the causal lines between our work and the stories that appear—but my guess is that our increased efforts have led to more coverage of GiveWell and our top charities this giving season than in the recent past.
• Website improvements. Companies that sell products online invest significant effort into optimizing their websites and checkout pages to maximize their revenues. We retained a marketing consultant, Will Wong of Mission Street, and we’ve been A/B testing different donation pages and plan to test other pages on our website such as our homepage or top charities page to see whether we can increase our conversion rate (i.e., the percentage of visitors to our website who give to one of our top charities). For context, our current conversion rate is 1%. Our understanding is that a standard conversion rate for e-commerce companies is 2%, and that international nonprofits have a similar conversion rate.5See Pg 51 of the study downloadable here. jQuery("#footnote_plugin_tooltip_5").tooltip({ tip: "#footnote_plugin_tooltip_text_5", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); An increase in our conversion rate to the industry average would lead to a significant increase in the amount of money we direct to our top charities.

Notes   [ + ]

1. ↑ We’ve also been running ads on Julia Galef’s Rationally Speaking podcast since then. Because it’s much smaller and more targeted, we’ve excluded it from this analysis. Measured returns to advertising on Rationally Speaking have been significantly better than the more mainstream podcasts discussed in this post. 2. ↑ I say “measure” retention because we’ve learned that many donors give subsequent donations directly to our top charities and don’t report those donations to us. We’ve tried to follow up with lapsed donors and with charities to track these donors down. 3. ↑ We only projected donations over five years. This is fairly arbitrary because we don’t have long-term enough data to know whether or not this is a reasonable assumption. We capped it to prevent our assessment being driven by speculation about how much money would be donated many years in the future. 4. ↑ I took this rough estimate from footnote 26, on page 15, of GiveWell’s 2015 metrics report. 5. ↑ See Pg 51 of the study downloadable here. function footnote_expand_reference_container() { jQuery("#footnote_references_container").show(); jQuery("#footnote_reference_container_collapse_button").text("-"); } function footnote_collapse_reference_container() { jQuery("#footnote_references_container").hide(); jQuery("#footnote_reference_container_collapse_button").text("+"); } function footnote_expand_collapse_reference_container() { if (jQuery("#footnote_references_container").is(":hidden")) { footnote_expand_reference_container(); } else { footnote_collapse_reference_container(); } } function footnote_moveToAnchor(p_str_TargetID) { footnote_expand_reference_container(); var l_obj_Target = jQuery("#" + p_str_TargetID); if(l_obj_Target.length) { jQuery('html, body').animate({ scrollTop: l_obj_Target.offset().top - window.innerHeight/2 }, 1000); } }

The post Update on our work on outreach appeared first on The GiveWell Blog.

GiveWell’s impact is a function of the quality of our research and the amount of money we direct to our recommended charities (our “money moved”). Historically, we’ve focused mostly on research because we felt that the quality of our recommendations was a greater constraint to our impact than our money moved.

This has changed. Outreach is now a major organizational priority. The goal of this work is to increase the amount of money we direct to our top-recommended charities.

In April 2014 I wrote about our work on outreach to explain why we hadn’t prioritized it: in brief, our growth had largely been driven by inbound interest in GiveWell, and proactive outreach efforts (beyond building relationships with existing donors) hadn’t yielded results that were worth the cost.

What changed?

• We believe that the amount of money we move is now a greater constraint to our impact than additional improvements in the quality of our research. Over the last two years, we’ve added five new top charities (three of which implement programs that weren’t previously represented on our top charities list), and we expect that our top charities, collectively, will have more than $200 million in unfilled funding gaps once they’ve received the funding that we expect to direct to them. (This calculation excludes GiveDirectly, which we believe could absorb and distribute hundreds of millions of dollars.) At the same time, the quality of our research and our capacity for research is higher than it’s ever been, so the returns to adding staff there (in terms of the pace at which we identify significantly better giving opportunities) are now lower.
• Increased capacity for outreach. In our 2014 post, we wrote that one of our key constraints was that senior staff (which at the time meant primarily GiveWell Co-Founder Holden Karnofsky and me) were necessary for most outreach-related work. This has changed. We now have capacity to take on outreach work as other staff have been hired and trained on this type of work.
• Better information on the impact of GiveWell’s outreach. We now have better information about the returns to outreach because:
  1. We’ve collected better data (via an improved donations processing system and outreach efforts) about where donors find out about us. Because of our ability to track donors, we know that a single appearance on NPR or major podcasts tends to drive $50,000+ in annual donations.
  2. More time passing has demonstrated that the lifetime value of the donations of a first time donor is higher than we expected. In several cases, we’ve seen major donors (i.e., those giving $10,000-$100,000) increase their annual giving by a factor of 10 or more.

We’re in the early stages of figuring out how we can proactively invest time and money in outreach to significantly increase our money moved. For now, we’ve taken some opportunities that we think will have positive returns; these are the three that we’ve invested the most time and money in to date:

• Podcast advertising. We’ve been advertising on podcasts that we believe our target audience listens to, based on interviews with current donors and GiveWell staff. In February and March, we ran a small experiment with a few ads on FiveThirtyEight’s Politics podcast and Vox’s The Weeds.1We’ve also been running ads on Julia Galef’s Rationally Speaking podcast since then. Because it’s much smaller and more targeted, we’ve excluded it from this analysis. Measured returns to advertising on Rationally Speaking have been significantly better than the more mainstream podcasts discussed in this post. jQuery("#footnote_plugin_tooltip_1").tooltip({ tip: "#footnote_plugin_tooltip_text_1", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

  In total, we spent approximately $20,000 on ads for this initial experiment. We ask donors who give via our website to tell us where they learned about GiveWell when they donate. GiveWell received approximately $8,000 in donations between February 1 and November 20 from donors who reported that they had learned about us via these podcasts.

  The donations we received were from first-time donors; to assess the impact of our advertising, we need to estimate the lifetime value of acquiring a new donor. In work we’ve done to assess our retention rate, we’ve seen that (a) approximately 20-25% of the donors who make a first-time donation of less than $1,000 give again in the subsequent year but (b) because many first-time donors increase the size of their donation over time, collectively, the donors who recur give more than 100% of the value of what they give in their first year.

  At higher donation levels ($1,000-$100,000), we measure 40-45% retention among donors, which leads to retention of approximately two-thirds of dollars given.2I say “measure” retention because we’ve learned that many donors give subsequent donations directly to our top charities and don’t report those donations to us. We’ve tried to follow up with lapsed donors and with charities to track these donors down. jQuery("#footnote_plugin_tooltip_2").tooltip({ tip: "#footnote_plugin_tooltip_text_2", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

  We therefore estimated the net present value of expected future donations (over the next five years) from these podcasts ads as somewhere between approximately $20,000 (assuming two-thirds dollar retention for the first two years and 100% dollar retention subsequently) and $45,000 (assuming 100% dollar retention).3We only projected donations over five years. This is fairly arbitrary because we don’t have long-term enough data to know whether or not this is a reasonable assumption. We capped it to prevent our assessment being driven by speculation about how much money would be donated many years in the future. jQuery("#footnote_plugin_tooltip_3").tooltip({ tip: "#footnote_plugin_tooltip_text_3", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });

  A few additional facts are worth keeping in mind about the above figures:

  • We ran this experiment in February and March; most donors give at the end of the calendar year. We consistently see donors who find out about GiveWell during the course of the year, but donate in December. Other things equal, we expect that our advertising would have had greater measured returns in December than earlier in the year.
  • We are only able to track donors who (a) fill out our donation form telling us where they learned about us and (b) give directly through our website rather than to our top charities. Less than 50% of donors who give via credit card (and a smaller percentage of donors who give via check) tell us where they learned about GiveWell. Also, roughly speaking, approximately 50% of the donors and dollars we influence come through GiveWell rather than going to our top charities.4I took this rough estimate from footnote 26, on page 15, of GiveWell’s 2015 metrics report. jQuery("#footnote_plugin_tooltip_4").tooltip({ tip: "#footnote_plugin_tooltip_text_4", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] });
  • It’s certainly possible that donors who learn about us via podcast would be more likely to give through our website than an average donor, more likely to report on how they found us (since their source is clear), or less likely to be retained. My best guess is that donors who learn about us via podcast ads behave similarly to our other donors, but I won’t be surprised if they don’t.

  With all that in mind, I believe that the impact of our podcast advertising is higher than what we directly measured.

  The results we saw from February to November this year were promising enough that we decided to increase the size of our experiment by spending approximately $100,000 on podcast ads. We’re currently running ads on FiveThirtyEight’s Politics podcast and Ezra Klein’s podcast and The Weeds at Vox.

• Earned media outreach. Mentions of GiveWell in the media have historically been a strong driver of growth. We aimed to increase mentions of GiveWell in high-quality, high-profile media where we’ve had the most past success as measured by dollars donated (i.e., media like The New York Times, NPR, The Wall Street Journal, and Financial Times). We retained a PR firm that came strongly recommended; we also increased 1-to-1 outreach by GiveWell staff to members of the media who have covered GiveWell in the past. It’s very hard to attribute the impact of the additional effort we’ve invested—overall, our effort has been fairly limited, and it’s hard to easily draw the causal lines between our work and the stories that appear—but my guess is that our increased efforts have led to more coverage of GiveWell and our top charities this giving season than in the recent past.
• Website improvements. Companies that sell products online invest significant effort into optimizing their websites and checkout pages to maximize their revenues. We retained a marketing consultant, Will Wong of Mission Street, and we’ve been A/B testing different donation pages and plan to test other pages on our website such as our homepage or top charities page to see whether we can increase our conversion rate (i.e., the percentage of visitors to our website who give to one of our top charities). For context, our current conversion rate is 1%. Our understanding is that a standard conversion rate for e-commerce companies is 2%, and that international nonprofits have a similar conversion rate.5See Pg 51 of the study downloadable here. jQuery("#footnote_plugin_tooltip_5").tooltip({ tip: "#footnote_plugin_tooltip_text_5", tipClass: "footnote_tooltip", effect: "fade", fadeOutSpeed: 100, predelay: 400, position: "top right", relative: true, offset: [10, 10] }); An increase in our conversion rate to the industry average would lead to a significant increase in the amount of money we direct to our top charities.

Notes   [ + ]

1. ↑ We’ve also been running ads on Julia Galef’s Rationally Speaking podcast since then. Because it’s much smaller and more targeted, we’ve excluded it from this analysis. Measured returns to advertising on Rationally Speaking have been significantly better than the more mainstream podcasts discussed in this post. 2. ↑ I say “measure” retention because we’ve learned that many donors give subsequent donations directly to our top charities and don’t report those donations to us. We’ve tried to follow up with lapsed donors and with charities to track these donors down. 3. ↑ We only projected donations over five years. This is fairly arbitrary because we don’t have long-term enough data to know whether or not this is a reasonable assumption. We capped it to prevent our assessment being driven by speculation about how much money would be donated many years in the future. 4. ↑ I took this rough estimate from footnote 26, on page 15, of GiveWell’s 2015 metrics report. 5. ↑ See Pg 51 of the study downloadable here. function footnote_expand_reference_container() { jQuery("#footnote_references_container").show(); jQuery("#footnote_reference_container_collapse_button").text("-"); } function footnote_collapse_reference_container() { jQuery("#footnote_references_container").hide(); jQuery("#footnote_reference_container_collapse_button").text("+"); } function footnote_expand_collapse_reference_container() { if (jQuery("#footnote_references_container").is(":hidden")) { footnote_expand_reference_container(); } else { footnote_collapse_reference_container(); } } function footnote_moveToAnchor(p_str_TargetID) { footnote_expand_reference_container(); var l_obj_Target = jQuery("#" + p_str_TargetID); if(l_obj_Target.length) { jQuery('html, body').animate({ scrollTop: l_obj_Target.offset().top - window.innerHeight/2 }, 1000); } }

The post Update on our work on outreach appeared first on The GiveWell Blog.

We recently discussed how you can give to reduce the administrative burden on charities. This post will focus on how you can save money on fees and give tax-efficiently so that more of your charitable budget can go directly to the organizations you want to support. This is an updated version of a post we originally ran in 2012; some content is the same, other content has been added or updated.

1. Don’t wait until the last minute. Many donors wait until the very end of the calendar year to give. If you’re hoping to make a donation by that deadline, we strongly advise against this. Here’s why:
   • Some methods of donating require some planning and preparation, such as giving appreciated stock.
   • Checks are tax-deductible according to the postmarked date on the envelope—you can’t write a check in 2018, backdate it to 2017, and claim a deduction. So, please head to the post office before the new year if you’re looking to get a tax deduction this year.
   • Leaving little time between making your donation and the deadline means you’ll have limited time to react if something unexpected happens, such as your credit card charge being declined.

   We recommend building in a cushion of a week or two if you’re aiming to donate by a particular deadline. The earlier you can give, the less likely you’ll have any issues. For end-of-year giving, we recommend a target date of December 24 or earlier.

2. Try to get a tax benefit. Details vary by country and personal situation, but a tax deduction can allow you to give much more to charity at the same cost to yourself. (That said, as discussed later in the post, we believe it is more important to give to the most effective possible charity than to get the maximum tax benefit.) Below, we discuss our understanding of donation methods for tax-advantaged giving, although please note that none of this information should be construed as legal or tax advice.

   Donors in the United States may make tax-deductible gifts to any of our nine top charities by giving to GiveWell. There are also a large number of tax-deductible options for giving to our top charities in other countries; please see the table here for more information.

   Donors in certain U.S. states and income brackets who are interested in maximizing their tax deduction may also consider “donation bunching,” or making two donations in one year rather than one donation in each of two years to take advantage of the standard deduction in one year and maximize the size of their itemized charitable deduction in a subsequent year. Considerations related to donation bunching are discussed in this post by former GiveWell intern Ben Kuhn.

3. Avoid the large transaction fees and delays associated with large online donations. When donating via credit card, you will almost always be charged standard credit card processing fees. Making a large donation via credit card may also trigger your card’s fraud protection (though a call to the credit card company can generally resolve the situation quickly).

   We discussed some of the tradeoffs between the ease of donating via certain platforms and the fees for donors and the administrative costs to charities for processing them in a previous post. In short, we do not advise making donations via credit card if you’re planning to give $1,000 or more.

4. Give appreciated stock and cryptocurrency. In the U.S., if you give stock or cryptocurrency (such as Bitcoin) to a charity, neither you nor the charity will have to pay taxes on capital gains (as you would if you sold the stock yourself). If you have stock or cryptocurrency that you acquired for $1,000 (and has a cost basis of $1,000) but is now worth $2,000, you can give the stock to charity, take a deduction for $2,000, and not have to pay capital gains tax on the $1,000 of appreciation. This can result in significant savings.

   Due to the administrative cost associated with processing donations of stock, we ask that donors donate stock directly to GiveWell only if the value of the stock at the time of transfer is estimated at approximately $1,000 or more. More information on giving appreciated stock to GiveWell, either through E*Trade or GiveWell’s Vanguard donor-advised fund, is available here. You can also use Betterment to donate appreciated stock to GiveWell. If you’re interested in making a Bitcoin donation to GiveWell, please email us at donations@givewell.org to receive instructions on how to give.

5. Look into donor-advised funds to make the process smoother and more consistent year-to-year. Donor-advised funds allow donors to make a charitable donation (and get a tax deduction) now, while deciding which charity they’d like to support later. The donation goes into a fund that is “advised” by the donor, and the donor may later recommend a grant from the fund to the charity of his/her choice.

   We see a couple of advantages to this setup. One advantage is that you can separate your “decision date” (the date on which you decide which charity you’d like to support) from your “transaction date.” That means that if you aren’t ready to decide which charity to support yet, you can still get started on the process of transferring funds and getting a tax deduction for the appropriate year. Another advantage is that if you change the charity you support from year to year, you’re still working with the same partner when it comes to transactions, so the process for e.g. donating stock will not change from year to year. Donor-advised funds are often set up to easily accept donated stock or non-traditional assets, whereas charities may or may not be.

   Many large investment companies—Vanguard, Fidelity, Schwab—offer donor-advised funds. They generally charge relatively modest management fees. We also maintain our own donor-advised fund for donors interested in supporting our recommended charities; the minimum size for a donation is $5,000. The GiveWell donor-advised fund is likely most helpful for donors interested in giving certain types of securities, such as Vanguard mutual funds, that are not accepted by E*Trade.

6. Find out if your company offers donation matching. Many companies offer to match employees’ gifts up to a certain amount. We recommend checking with your employer if you’re unsure whether they offer this option. Some employers have a limited list of charities to which they will match donations; consider asking your employer whether they would add the charity of your choosing if it isn’t already on the list.
7. Consider the political environment. If you believe that your likelihood of taking charitable deductions is higher in 2017 than it will be in 2018, consider increasing your giving this year.
8. Choose your charity wisely. Saving money on taxes and transaction fees can be significant, in some cases approaching or exceeding a 50 percent increase in the amount you’re able to give. However, we believe that your choice of charity is a much larger factor in how much good your giving accomplishes.

   Our charity recommendations make it possible to support outstanding, thoroughly vetted organizations—which we’ve investigated by reviewing academic evidence, interviewing staff, analyzing internal documents, conducting site visits, assessing funding needs, and more—without needing to do your own research. We publicly publish the full details of our process and analysis, so you can spot-check whatever part of our work and reasoning you’d like to.

Final notes

If you support our recommended charities (on the basis of our recommendation) but you don’t give through our website, please fill out this form to let us know about your gift; doing so helps GiveWell track our impact.

We believe that even when dealing with a relatively complicated gift (for example, a gift of stock), it’s possible to give quite quickly and with only minor hassle. The much more difficult challenge is choosing a charity, and we’ve tried to make that easy as well. We hope you’ll give this season, even if you’re just starting to think about it now.

If you’d like more advice about how to donate, please don’t hesitate to contact us. We’re happy to talk.

The post Maximizing the impact of your donation: saving on fees means more money for great charities appeared first on The GiveWell Blog.