You are here

Approaches to Moral Weights: How GiveWell Compares to Other Actors

Published: November 2017

In a previous blog post, we described how we use cost-effectiveness analyses when deciding which charities to recommend to donors.

This report discusses how GiveWell and other actors, such as governments and global health organizations, approach one of the most subjective and uncertain inputs into cost-effectiveness analyses: how to morally value different good outcomes.

For example, GiveDirectly, one of GiveWell's seven top charities, increases recipients' consumption, while the primary benefit we see from our top charity the Against Malaria Foundation is that it averts the deaths of young children. How can one make a direct comparison between the amount of "good" achieved by each of these charities?

GiveWell does this by assigning quantitative "moral weights" to different outcomes in our cost-effectiveness analyses. As a check on how sensitive our recommendations are to our moral assumptions, we investigated how others typically answer these questions in their cost-effectiveness analyses. This report discusses our findings from this investigation.

Summary

We focus on the following questions:

  • Why does GiveWell explicitly include moral weights in our cost-effectiveness analyses, and how do we decide on moral weights?
  • Is there a "standard" approach to moral weights in cost-effectiveness analyses? How do other actors, such as governments and the World Health Organization, make these judgments?
  • How much would GiveWell's cost-effectiveness analyses change if we took a "standard" approach to moral weights?

In brief:

  • We include moral weights in our cost-effectiveness analyses because they are an important part of any giving decision and we think it is valuable to be transparent about them. The moral weights that drive our cost-effectiveness estimates are based on our staff's personal values.1
  • Governments and other prominent actors often use "value of a statistical life" estimates to compare the value of improving health relative to raising incomes. These estimates often imply that a year of healthy life is roughly 2-3x as valuable as a year of doubling someone's income. However, there is little relevant research to inform such estimates in low- and middle-income country (LMIC) contexts; we understand that how income is valued relative to health may shift when a population is much poorer.
  • There does not seem to be a standard approach for comparing the value of life at different ages; the most commonly used framework that we have seen (the disability-adjusted life year framework) explicitly does not provide judgments on this topic. Nevertheless, most other analyses that we have seen assume that averting death during childhood is about 1-2x more valuable than averting death during adulthood.
  • Our initial analysis suggests that using relatively "standard" moral weight assumptions (i.e., the assumptions in the previous two bullet points) instead of our staff's moral weights would not change our overall view of the relative cost-effectiveness of our current top charities. It may affect how we view some interventions in the future, particularly those that disproportionately focus on averting deaths for young children or adults. We plan to include explicit comparisons between staff moral weights and relatively "standard" moral weights in our analyses going forward.

Why does GiveWell explicitly include moral weights in its cost-effectiveness analyses, and how does GiveWell decide on moral weights?

Why we include moral weights

GiveWell aims to find and recommend charities that are evidence-backed and cost-effective and to make a bottom-line recommendation for donors about where to give. In order to do this, we spend a lot of time thinking about the relative cost-effectiveness of different opportunities.

In order to reach an overall comparison of the cost-effectiveness of different charities, we first use empirical analysis to estimate the cost per outcome achieved by different programs. For example, we currently estimate that it costs about $7,000 to avert the death of one under-5-year-old by distributing malaria nets via the Against Malaria Foundation and that it costs about $1,200 to double a household's consumption for a year via GiveDirectly.2 These estimates are subject to substantial uncertainty and should not be taken literally, but they largely involve judgment calls about evidence—for example, how similar you think the organization's current work is to the randomized controlled trials that were done of its program—and not ethics.

However, when it comes to making a decision about where to donate or how cost-effective one charity is relative to another, one also needs to consider moral questions such as:

  • Valuing health vs. income: How much do I value averting the death of a 2-year-old relative to doubling the income of an extremely poor household?
  • "Age-weighting": How much do I value averting the death of a 2-year-old relative to averting the death of a 30-year-old?

Setting moral weights (incorporating the value judgments in the bullet points above, or similar, by giving them numeric value in our cost-effectiveness analyses) is uncomfortable and challenging. In response to a question about whether giving insecticide-treated nets (bed nets) to prevent malaria or giving cash transfers would be a better use of marginal funds, the development economist Esther Duflo responded:3

I think it’s the type of question that will be frankly difficult to address because you’re required to value benefits across sectors. So it depends what your objective is. If your objective is to control malaria, you can do an experiment where you give money and you give people bed nets or you do something else and you can see how many people sleep under a bed net at the end of the day, and what’s the number of malaria infections you’ve averted. So that’s a question that’s well defined. Whether it’s better to give cash or give bed nets would require you to make a judgement about what is the importance of making people healthy versus having them buy a roof. And that’s one I’m not prepared to make.

However, answering these questions is unavoidable. Anyone deciding to donate to one charity over another is implicitly using moral weights, even if that person is not explicitly engaging with them.

GiveWell openly engages with these questions for a variety of reasons, including:

  • We want to be transparent about the moral values underlying our recommendations,
  • We want to give donors who rely on our research the opportunity to change the moral weights in our analyses so that they can choose a charity based on their own values, and
  • We hope that by being explicit about these tradeoffs, we will be more likely to reflect carefully on them and find the best giving opportunities according to our values.

How we include moral weights

We would ideally incorporate the views of additional people who don't work on GiveWell directly, including our recommended charities' beneficiaries, in the moral weights we use. Unfortunately, little information about others' moral weights exists. We are partially working to address this by funding new research on beneficiaries' preferences through our Incubation Grants program.

In the meantime, we set moral weights by asking our staff for their values; when setting their weights, staff consider a variety of factors including the approaches of other organizations (described below). You can see the moral weights that our staff assign to different outcomes here. Anyone can make a copy of our cost-effectiveness analysis from this page and input their own moral weights to determine which charity is most cost-effective, given their values.

For previous discussion of some philosophical considerations relevant to setting moral weights, see our December 2016 blog post on this topic.

Is there a "standard" approach to moral weights in cost-effectiveness analysis?

We found that:

  • Using estimates of the "value of a statistical life" seems to be a fairly standard approach by governments and other major actors to compare the value of income relative to health. These estimates typically find that one year of healthy life is worth about two to three times a country's gross domestic product per capita, though there is high variability in estimates and major methodological limitations of the research on which they are based. In addition, little of this research has been conducted in LMIC contexts, where GiveWell's top charities work, and so may have limited applicability to outcomes there.
  • There is less relevant literature and discussion about the value of averting deaths at different ages than there is about estimating the value of a statistical life. The most common approach that we have seen for valuing averting death at different ages is the disability-adjusted life year (DALY) framework. This framework is not intended to fully account for moral considerations. Our impression is that there is not a "standard" way to assign moral weights related to age. However, most other analyses that we have seen typically assume that the lives of children are about one to two times as valuable as the lives of adults.

Research process

In order to assess the sensitivity of our recommendations to our staff's moral weights, we looked at how others approach these questions. To limit this investigation, we began by focusing on the assumptions other actors make about the two questions mentioned above: 1) how to value income relative to health, and 2) how to value averting the deaths of young children relative to those of adults. We focused particularly on the approaches of governments and international institutions, such as the World Health Organization, because these actors play major roles in allocating resources in a variety of contexts.

The below represents our impressions based on a literature review and talking to researchers who have worked on these topics.4 In general, the researchers that we have spoken with seem to agree that there is too little research on this topic given its importance.

Valuing income versus health

Our understanding is that the most common way to estimate the value of income versus health is to estimate how much people are willing to pay to avert death or to add healthy years to their lives. Such estimates are often presented as the "value of a statistical life" (VSL) or the cost per disability-adjusted life-year (DALY) averted.

Governments and other major international actors such as the World Health Organization sometimes use such estimates to determine which programs to support. For example, if a government is considering an environmental regulation that would decrease economic output but also save many lives, it might estimate the value of a life to determine whether the benefits of the regulation outweigh the costs. Or say, for example, that an international organization distributes medicine and cash. If we assume that someone would be willing to pay $3,000 for an additional year of healthy life and that it would cost $1,000 to purchase a medicine that would provide them with an additional year of healthy life, then it is 3x more cost-effective to provide that person with the medicine than to give them $3,000, or the amount of cash that would achieve an equivalently good outcome.5

Revealed and stated preference research

Governments and other actors generally use at least two major methodologies to arrive at these estimates: "revealed preference" research and "stated preference" research. Revealed preference methods look at people's choices in real-world environments to assess how much they must be paid to take a particular risk of death. For example, they may estimate how much additional money someone needs to be paid to take a job that carries a 1% higher mortality risk than similar jobs they could attain. Stated preference methods directly ask people questions about these tradeoffs, such as how much they would be willing to pay to reduce their risk of death by 1 in 10,000.6

The U.S. government more commonly uses revealed preference analyses to estimate VSL; other Organisation for Economic Co-operation and Development (OECD) countries more commonly use stated preference analyses. We do not know why these groups prefer different methodologies.7

Research based on these methods often concludes that a year of healthy life is roughly as valuable as 2-3x gross domestic product (GDP) per capita. For example:

  • WHO’s CHOosing Interventions that are Cost-Effective (CHOICE) team, which assists country policymakers with decisionmaking, distinguishes between the following tiers of cost-effectiveness:8
    • "Very cost-effective": Cost per DALY averted is less than GDP per capita
    • "Cost-effective": Cost per DALY averted is between 1-3x GDP per capita
    • "Not cost-effective": Cost per DALY averted is greater than 3x GDP per capita
  • The Lancet Commission on Investing in Health's "Global Health 2035" project estimated "that the value of a life year (VLY) averages 2.3 times GDP per capita for low and middle–income countries (LMICs)" based on U.S. VSL estimates that were adjusted for lower-income contexts using a variety of assumptions.9
  • High-income country governments often use VSLs that range from about $3 million to $7 million, which can be converted to a value per DALY of about $60,000 to $230,000, or about 1-6x GDP per capita.10

We have not yet vetted the research that leads to these estimates and we see a variety of major limitations of both revealed preference and stated preference research.

Limitations of these analyses

A few general issues that limit the usefulness of both types of research are:

  • Difficulty of comprehending small probabilities: It may be challenging for people to understand probabilities well enough to indicate how much they value small chances of averting death.
  • Lack of information: People may not have or consider basic information relevant to thinking about the value of mortality risks.
  • Preferences may not maximize well-being: Even if people perfectly understood the probability and information components of trading off income and mortality risk, they might not be able to reliably anticipate what would maximize their well-being, all things considered. This may apply to people in general, including both the populations surveyed in the literature mentioned above and the beneficiaries of programs recommended by GiveWell. For example, maybe people do not realize how important health is to their happiness and their long-term goals and undervalue it. Donors who use GiveWell's recommendations may want to consider the potential disconnect between preferences and well-being when making giving decisions.

A further challenge is that there is little revealed preference or stated preference research conducted in LMICs; most VSL and similar analyses estimate how much people value life in LMICs by extrapolating from high-income country research.11 A key issue with extrapolation is that one needs to make an assumption about how much the relative value of income versus health changes when a population is much poorer (often referred to as "the elasticity of demand for health"). Perhaps someone who barely has enough money to survive would greatly prefer any increase in income more than an additional year of life. Different assumptions about how to extrapolate can lead to estimates of the value of a DALY that vary by at least an order of magnitude.12

Though the literature on VSL in LMIC contexts is limited, we are aware of a few potentially relevant empirical papers on the topic, which are briefly summarized in León and Miguel 2016, itself an estimate of VSL in an LMIC context (see following footnote).13 These papers generally appear to find substantially lower values of health relative to income than are estimated in high-income countries.14 We have not yet carefully vetted these papers and expect to review them more closely in the future, but our impression is that estimates of the value of life from these papers have not yet been used by major decisionmakers and are based on different methodologies than typical VSL estimates, so they should not yet be interpreted as "standard" assumptions.15

Because of limitations in the existing literature, we do not see current "best guess" estimates of the relative value of income versus health in LMICs as robust.

Valuing deaths of young children versus adults

DALY framework

The method that we have seen used most often for valuing deaths averted at different ages, the DALY framework, is described and updated as part of the series of Global Burden of Disease (GBD) reports. GBD is a global observation epidemiological study that reports on mortality and morbidity from a variety of causes. The DALY framework is used by the World Health Organization (WHO)16 and the Disease Control Priorities Project (a major research initiative to help prioritize spending within global health),17 and our impression is that it is widely used in other research about the cost-effectiveness of global health and development programs.

Broadly, the DALY framework values a death averted by adding up the total years of life saved by an intervention. For example, the death of a male infant (life expectancy 80 years) would be counted as 80 years of life lost, while the death of a 45-year-old female (life expectancy 83 years) would be counted as 38 years of life lost. Without further adjustments, this implies that the death of a single infant is considered about as bad as the death of two adults. (We provide more background on the DALY framework in two 2008 blog posts: here and here.)

The DALY framework also provides optional adjustments to 1) factor in the idea that life years might be more valuable when one is in the "prime" of one's life (called "age-weighting"), and 2) to assign less value to years of life saved that occur farther in the future ("time discounting").18 The GBD has varied its assumptions about age-weighting and time discounting over its history.19 For example, the GBD's 2004 report included age-weighting and time discounting in its main analyses, but in 2010 both the GBD and WHO decided not to factor in either adjustment in their core analyses.20

The GBD and WHO's justifications for removing age-weighting and time discounting was that they wanted to leave these kinds of moral judgments to policymakers. Therefore, we do not interpret their decision as reflecting researchers' or decisionmakers' moral views.21 However, in practice we have not seen policymakers explicitly adjust the GBD's output based on their views about age weights and discount rates. This leaves us in a position where we are not able to look to one of the most commonly used tools in cost-effectiveness analysis (the DALY framework) for guidance on how to approach age-weighting and time discounting as part of our decisionmaking process.

Other estimates

In our brief literature search, we also came across a few other estimates of the relative value of averting deaths of young children versus adults, including:

  • An OECD literature review of VSL research that recommended, based on information about how much parents value averting mortality risk for their children, the VSL for a child should be ~1.5-2x higher than the mean adult VSL.22 We have not vetted the underlying methodology for this estimate and do not know whether any decisionmakers actually followed the OECD report's recommendation. The OECD literature review also noted that policymakers in the U.S. have been reluctant to make age adjustments to VSL figures because there was controversy when it was reported that the Environmental Protection Agency (EPA) assigned lower VSL estimates for elderly people as part of its analyses.23
  • A survey by Julian Jamison conducted via Amazon Mechanical Turk, a website on which workers can be paid to complete "human intelligence tasks." Jamison published a paper reporting that respondents valued adult women roughly 1.5-2x as much as fetuses and infants up to 1 week old, but that they valued 1-year-old children roughly equally to adult women.24 Jamison notes that roughly 40% of respondents failed to input a response and that this study was done on a relatively narrow population, among other limitations.25

Unfortunately, we did not see other highly relevant research on this topic in our preliminary search, though it is possible that a more comprehensive search would identify additional research.

Philosophers' views on moral weights

One might expect that questions such as these might be frequently researched and discussed by philosophers. However, even though philosophers have long considered the question of what makes life valuable in general, our understanding from speaking with experts and searching for literature is that philosophers have not done much work to consider how best to assign quantitative value to different kinds of outcomes.26 For further discussion of philosophical work that is potentially relevant to assigning moral weights, see our previous blog post on the Against Malaria Foundation and population ethics.

How much would GiveWell's cost-effectiveness analyses change if we took a "standard" approach to moral weights?

Changing our cost-effectiveness analyses to use "standard" moral weights rather than staff values would not substantially change the estimated relative cost-effectiveness of our current top charities (which mainly focus on averting children's deaths and increasing adult income and consumption), though it could make a large difference to our estimates of the impact of programs we may work on in the future, such as antiretroviral therapy (which is mainly focused on extending adult lives).

A table comparing the median cost-effectiveness estimates for top charities and antiretroviral therapy charities under varying assumptions is below:

Median cost-effectiveness, relative to unconditional cash transfers (GiveDirectly)27
Deworm the World Initiative Schistosomiasis Control Initiative Against Malaria Foundation Malaria Consortium Antiretroviral therapy
"Standard" moral weights ~10x ~8x ~4x ~5x ~0.5x
Actual staff moral weights ~10x ~8x ~3x ~3x ~2x

The above "standard" results are based on a version of our cost-effectiveness analysis that replaces staff moral weights with relatively "standard" moral weights based on the conclusions of the research discussed above. The "standard" moral weight assumptions that we made were:28

  • Averting a DALY is roughly as valuable as providing a cash transfer of ~2.5x GDP per capita.
  • Averting the death of a young child is equivalent to averting about 37 DALYs. This is consistent with the 2004 GBD methodology, which factored in discounting and age-weighting.29 We are uncertain whether this estimate was intended to fully reflect the GBD researchers' moral views at the time, but it seems to us more likely to be a reflection of their moral views than the 2010 GBD figures, which exclude age-weighting and time discounting adjustments.
  • Averting the death of an adult is equivalent to averting about 30 DALYs. This also relies on the 2004 GBD methodology.

All staff-specific parameters for judgment calls about evidence (e.g., how much to discount the expected effect of deworming due to concerns about evidence quality) have not been changed in this model.

It seems that using "standard" assumptions could make a large difference to our cost-effectiveness estimates for programs that are primarily focused on reducing mortality for very young children or mortality for adults, as illustrated by the relatively large cost-effectiveness changes under different assumptions for Malaria Consortium's seasonal malaria chemoprevention program (mainly focused on averting infant deaths) and antiretroviral therapy (mainly focused on extending adult lives).30 GiveWell staff often value averting adult deaths about 1-5x more than infant deaths relative to "standard" approaches, so changes to these weights become particularly relevant for interventions where mortality reduction of either group is the primary outcome.31

The cost-effectiveness of deworming charities and the Against Malaria Foundation did not change considerably under "standard" assumptions; further explanation for this is in the following footnote.32

In future cost-effectiveness analyses, we plan to include an entry for "standard" assumptions as one of the published inputs so that we and others who use our research can track how much our moral weights differ from our best guess of the kinds of inputs that would be used by other mainstream decisionmakers.

We may also experiment with surveying policymakers and other researchers about their moral weights, which could then be used in our model. However, in our brief, initial efforts, we have struggled to find individuals who are willing to provide their views on these topics.

Sources

Document Source
Chang et al. 2017 Source (archive)
Deaton et al. 2009 Source (archive)
Disease Control Priorities in Developing Countries, "Chapter 2: Intervention Cost-Effectiveness: Overview of Main Messages" Source (archive)
GiveWell's non-verbatim summary of a conversation with Dean Jamison, January 9, 2017 Source
GiveWell's non-verbatim summary of a conversation with S. Andrew Schroeder, September 13, 2016 Source
Global Health 2035, "Supplementary web appendix 3: Valuation of changes in mortality rates" Source (archive)
Jamison 2016 Source (archive)
Kremer et al. 2011 Source (archive)
León and Miguel 2016 Source (archive)
Murray et al. 2010, "GBD 2010: design, definitions, and metrics" Source (archive)
OECD, "Mortality Risk Valuation in Environment, Health and Transport Policies." Source (archive)
Robinson et al. 2017 Source (archive)
WHO 2013, "WHO methods and data sources for global burden of disease estimates 2000-2011" Source (archive)
WHO, "Disability weights, discounting and age weighting of DALYs" Source (archive)
  • 1.

    Not all of the people who contribute moral values to our cost-effectiveness estimates are full-time GiveWell employees. For example, James Snowden contributes inputs as a GiveWell Research Consultant and Holden Karnofsky, now a GiveWell Board member, contributes inputs since he has had significant experience with our estimates as the former Co-Executive Director of GiveWell. We use the term "staff" in this report for simplicity.

  • 2.

    Note: This report primarily references figures from the May 2017 version of our cost-effectiveness analysis which has since been updated. Changes to our analysis may cause the numbers in this report to become outdated.

    • Estimates for the cost per under-5-year-old death averted by the Against Malaria Foundation and the cost to double a household's consumption for a year via GiveDirectly vary slightly by staff member.
    • For estimates of cost per under-5-year old-death averted by the Against Malaria Foundation, see Row 44, Sheet "Nets" in our August 2017 CEA.
    • For an estimate of cost to double a household's consumption for the equivalent of a year via GiveDirectly, see Cell B40, Sheet "Cash" in an edited version of our May 2017 CEA.
    • For more background on our cost-effectiveness analyses, see this page.
  • 3.

    See Timothy Ogden's interview with Esther Duflo in Experimental Conversations. Note that Odgen is a member of GiveWell's Board of Directors.

  • 4.

    Unfortunately, we were not able to publish notes from all conversations we had with researchers, but for two examples, see:

  • 5.

    It is worth noting that we have not seen other actors explicitly use this kind of analysis to decide between distributing cash transfers and supporting health interventions; it may be that they would conduct the analysis differently if they were facing this kind of decision.

  • 6.

    "There are two main methodological traditions to value mortality risk changes, and VSL, in monetary terms: revealed and stated preference methods. Revealed Preference (RP) methods are based on individual behaviour in markets where prices reflect differences in mortality risk (e.g. a labour market, where wages reflect differences in workplace mortality risks), and markets for products that reduce or eliminate mortality risks (e.g. buying bottled water to reduce mortality risk from contaminated tap or well water, and buying motorcycle helmets to reduce mortality risks in traffic accidents). These two RP approaches, termed the “hedonic wage” (HW)/wage risk (see e.g. Viscusi and Aldy, 2003) and “averting costs” (AC) methods (see e.g. Blomquist, 2004), respectively, depend on a set of strict assumptions about the market and the respondents’ information and behaviour which are seldom fulfilled.

    Stated Preference (SP) methods, e.g. contingent valuation (CV) or choice modelling (CM), instead construct a hypothetical market for the mortality risk change in question and ask respondents directly in surveys for their willingness-to-pay (WTP) to reduce their mortality risk, from which the VSL can then be derived. Both RP and SP methods have their strengths and weaknesses, but there has been a growing emphasis on SP methods in recent years. Important reasons for this is that many environmental, transport and health policies affect the youngest or the oldest part of the population the most (rather than the workers in occupations that involve risk, whom wage risk studies are based on), and that mortality often results from long-term risk exposure and exacerbation of pre-existing medical conditions (rather than accidental deaths in the workplace)." Pg. 19, OECD, "Mortality Risk Valuation in Environment, Health and Transport Policies."

  • 7.

    "However, the method used to establish a VSL number for policy making vary widely between countries, and even between agencies within a country. The main difference is the reliance on Revealed Preference (RP) methods in terms of wage risk studies in the United States (where most such studies have been conducted), while Europe, Canada and Australia rely more on Stated Preference (SP) methods, eliciting people’s willingness-to-pay (WTP) for changes in mortality risks." Pgs. 13-14, OECD, "Mortality Risk Valuation in Environment, Health and Transport Policies."

  • 8.
    • "For many years, cost-effectiveness thresholds of one and three times GDP per capita per DALY averted have been frequently cited in global health. For example, the World Health Organization’s (WHO’s) Choosing Interventions that are Cost-Effective (CHOICE) program defines interventions for which the cost per DALY averted is less than GDP per capita as very cost-effective, between one and three times GDP per capita as cost-effective, and greater than three times GDP per capita as not cost-effective." Robinson et al. 2017, Pg. 142.
    • Robinson et al. 2017 notes that these rough estimates of the value of a DALY seem to be based on a variety of sources, including U.S. VSL estimates, which could also support multipliers larger than three. This represents one example where U.S. VSL estimates seem to have been extrapolated to LMIC contexts.
      • "These are demand-based values taken from the 2001 report of WHO’s Commission on Macroeconomics and Health (CMH). The CMH does not explicitly address cost-effectiveness thresholds; rather, it develops estimates for use in benefit-cost analysis. The CMH notes, "According to some estimates, each life year is valued at about three times the annual earnings. This multiple of earnings reflects the value of leisure time in addition to market consumption, the pure longevity effect, and the pain and suffering associated with disease. (CMH 2001, p. 31)"...The CMH reports its analytic results using one times GNI per capita as a ‘very conservative’ estimate of the value of a life year and also applies the three times GNI per capita value per life year noted earlier. The CMH does not describe the derivation of these values, noting only that ‘[s]uch high valuations have been used in several recent economic analyses’ (p. 31). It references four examples. The first is Cutler and Richardson (1997), who apply a value per life year of $100 000 in 1990 dollars as their benchmark US value based on Tolley et al. (1994). This is somewhat less than the midpoint of the $70 000–$175 000 range Tolley et al. derive using various value per statistical life (VSL) estimates and discount rates. It is approximately four times 1990 US per capita GDP or GNI, which were both $24 000 at that time.4 The second reference is to a 1999 working paper by KM Murphy and R Topel (unpublished), who use a life cycle model that combines theoretical expectations with VSL research and data on US earnings, consumption and life expectancy at different ages to estimate the value of increasing life spans.5 They find that the present value of a 1 year change in life expectancy is $150 000–$200 000 in 1992 dollars. This range is approximately six to eight times 1992 US per capita GDP or GNI, which were somewhat above $25 000. The third and fourth references, G Becker et al. (unpublished) and Philipson and Soares (2001), are closely related.6 Both are working papers that focus on calculating full income for the purpose of cross-country comparisons, where full income includes both GDP per capita and the value of life expectancy. Each relies on a life cycle model to estimate the value per life year but neither provides a mean nor median value that can be compared with the values from the other studies.
        Thus the research cited by the CMH supports a value per life year greater than GNI or GDP per capita, perhaps by multipliers larger than three. The CMH multipliers were not rigorously derived; they were illustrative estimates based on the then-available research. The CMH is very clear that its calculations were intended as rough examples." Robinson et al. 2017, Pgs. 142-143.
  • 9.
    • Chang et al. 2017 summarize the Commission on Investing in Health as finding "that the value of a life year (VLY) averages 2.3 times GDP per capita for low and middle–income countries (LMICs)," but Chang et al. 2017 note that this estimate is highly sensitive to alternative assumptions. We have not yet vetted Chang et al. 2017's analysis.
    • The authors of "Global health 2035: a world converging within a generation" provide an explanation of their methodology for estimating the value of a life year in Global Health 2035, "Supplementary web appendix 3: Valuation of changes in mortality rates".
    • "In 'Global health 2035: a world converging within a generation,' The Lancet Commission on Investing in Health (CIH) adds the value of increased life expectancy to the value of growth in gross domestic product (GDP) when assessing national well-being. To value changes in life expectancy, the CIH relies on several strong assumptions to bridge gaps in the empirical research. It finds that the value of a life year (VLY) averages 2.3 times GDP per capita for low and middle–income countries (LMICs) assuming the changes in life expectancy they experienced from 2000 to 2011 are permanent." Chang et al. 2017, Abstract.
    • "We find that reasonable alternative assumptions regarding the effects of income, age, and life expectancy may reduce the VLY estimates to 0.2 to 2.1 times GDP per capita for LMICs. Removing the reduction for young children increases the VLY, while reversing the sequencing of the calculations reduces the VLY." Chang et al. 2017, Abstract.
    • "To support this approach, prominent members of the Commission, including economists Dean Jamison, Lawrence Summers, and Kenneth Arrow, developed an innovative method for estimating the value of an increase in population life expectancy and translating the results into an average value of a life year (VLY) [1]. They adjust a US estimate of the value of mortality risk reduction, assuming this value is proportional to GDP per capita and to remaining life expectancy. In their main result, they decrease the value for children aged 0 through 4 by 50 percent, then divide the total by the life expectancy gain to estimate VLY. This VLY is 2.3 times 2000 GDP per capita for life expectancy gains experienced by low– and middle–income countries (LMICs) from 2000 to 2011, assuming the gain is permanent, or 3.0 times GDP per capita without the reduction for young children. The CIH also reports the value of life expectancy gains for other country groups and time periods using the same approach, and applies the results to estimate the benefits of the interventions it recommends." Chang et al. 2017, Pgs. 1-2.
  • 10.
    • "The report summarises the results of a four-year effort to compile and analyse the largest database to date containing all SP studies that have been prepared around the world and that estimate adult VSL in environmental, health and transport risk contexts. The objective is to summarise this literature to answer two broad questions of relevance for both policy and research communities:
      1. What are the main factors explaining people’s WTP for reductions in mortality risks in the environmental, health and transport contexts, and the VSL derived from SP studies?
      2. Based on the current knowledge, which VSL estimates should be used in analysis of environmental, health and transport policies?

      ...The book proposes a range for the average adult VSL for OECD countries of USD (2005-USD) 1.5 million – 4.5 million, with a base value of USD 3 million. For EU-27, the corresponding range is USD 1.8 million – 5.4 million (2005-USD), with a base value of USD 3.6 million. These base values and ranges should be updated as new VSL primary studies are conducted." Executive summary, OECD, "Mortality Risk Valuation in Environment, Health and Transport Policies." Pgs. 14-15.

    • The above provides a range for a VSL of about $1.5 million to $5.5 million in 2005 USD. This range would be about $1.9 million to $7 million in 2017 USD (see here). Assuming that a VSL is equivalent to about 30 years of healthy life (see section on the DALY framework below for more details), this would be a range of roughly $60,000 to $230,000 per year of healthy life ($1.9 million / 30 = ~$63,000 and $7 million / 30 = ~$230,000). Assuming a GDP per capita of roughly $40K-$60K (see here for estimates of GDP per capita in the US and Europe), this would imply that a year of healthy life is roughly 1-6x as valuable as GDP per capita.
    • See Table 1.1, OECD, "Mortality Risk Valuation in Environment, Health and Transport Policies." Pg. 25. This table shows VSL estimates used by a variety of U.S. government agencies.
  • 11.

    Robinson et al. 2017 notes that most literature reviews on this topic that they are aware of are outdated, but that they expect that there have not been high quality studies on this topic completed in LMIC contexts: "A second important improvement would be to conduct a criteria-driven review of the VSL and VSLY studies globally. The available reviews cited earlier are outdated, and our understanding of what constitutes a high quality study is evolving as a result of new research and expert reviews. Evaluation criteria that reflect our current understanding of best practices are described in recent reviews of US VSL studies (Robinson and Hammitt 2016, U.S. Environmental Protection Agency 2016).11 These criteria can be adapted for application to studies conducted elsewhere, and also adjusted to address the available VSLY studies as well as any studies that explicitly explore the monetary value of a DALY. We expect that this review will indicate that extrapolation of values across countries continues to be necessary, but will aid in improving the approach for such extrapolation. It would provide insights into the effects of income differences on these values as well as the effects of other influencing factors." Robinson et al. 2017, Pg. 144.

  • 12.

    "In the near term, one important but simple improvement involves more explicitly accounting for uncertainties in the relationship between VSL and income by applying a range of income elasticity estimates. The CMH extrapolates from a US VSL estimate assuming an income elasticity of 1.0; as discussed in Hammitt and Robinson (2011), larger elasticities may be appropriate. Changing the elasticity can lead to values that vary by an order of magnitude or more. For example, starting with a US VSL of $9.0 million and using the GDP per capita estimates above, the Kenyan VSL would be $480 000 with an elasticity of 1.0, $110 000 with an elasticity of 1.5 and $25 000 with an elasticity of 2.0. If we assume that the remaining life span for the average Kenyan is 30.5 years, then the resulting VSLY would be $24 000, $5600 or $1300 depending on the elasticity (3% discount rate).10" Robinson et al. 2017, Pg. 144.

  • 13.
    • "The average VSL estimate for non-African travelers in our sample, who are typically from OECD countries, is US$924,000. This is somewhat lower than most previous rich country estimates, which typically use hedonic approaches and range from US$1 to 9.2 million, but is similar to Ashenfelter and Greenstone (2004b).[3: See, for example: Viscusi and Aldy (2003); Ashenfelter and Greenstone (2004b), Lee and Taylor (2012). Ashenfelter and Greenstone (2004a) argue that estimates in this literature are subject to an upward publication bias.]
      The only comparable estimates available from less developed countries (to our knowledge) are for manufacturing workers in India and Taiwan and range from US$0.5 to 1 million (Liu et al. 1997, Shanmugam 2001). These are in the same range as the estimates for the African travelers in our data, with an average VSL estimate of US$577,000 (PPP).[4: Previous studies in developing countries are based on compensating differentials in the labor market, which are often criticized for being particularly prone to selection bias. In the African context, Deaton et al. (2010) use a subjective life evaluation approach to estimate the monetary value attached to the death of a relative.] Kremer et al. (2011) use a travel cost approach – namely, willingness to walk longer distances to cleaner drinking water sources – to estimate the willingness to pay for avoiding a child diarrhea death among rural Kenyans, and find that it is very low in that setting, at under US$1,000." Pgs. 4-5, León and Miguel 2016.
    • "This paper exploits an unusual transportation setting to generate some of the first revealed preference value of a statistical life (VSL) estimates from a low-income setting. We estimate the trade-offs individuals are willing to make between mortality risk and cost as they travel to and from the international airport in Sierra Leone. The setting and original dataset allow us to address some typical omitted variable concerns, and also to compare VSL estimates for travelers from different countries, all facing the same choice situation. The average VSL estimate for African travelers in the sample is US$577,000 compared to US$924,000 for non-Africans." Abstract, León and Miguel 2016.
    • "Our estimates are based on the choices of middle-class and wealthy African travelers, and thus are complementary to Kremer et al.’s estimates derived from relatively poor rural households." Pg. 5, León and Miguel 2016.
  • 14.
    • "The fact that the estimated VSL for African travelers is somewhat lower than for non-African travelers (who are mainly from wealthy countries) is consistent with a growing body of research that documents the relatively low demand for health and life in less developed countries. The disease burden in low income countries is much higher than in rich countries, and yet a number of scholars have documented surprisingly low investments in preventive health technologies (Kremer and Miguel 2007; Kremer et al. 2011; Cohen and Dupas 2010; Dupas and Miguel 2016). Common explanations (surveyed in Dupas 2011) range from a lack of information about new health technologies (Madajewicz et al 2007), pervasive liquidity constraints (Tarozzi et al 2013), time inconsistent preferences (DellaVigna and Malmendier 2006), agency problems within the household (Ashraf et al. 2014), shorter life expectancy (Oster 2009), cultural attitudes (and especially fatalism, the belief that fate governs major life outcomes)5 and a high income elasticity of demand for health expenditures (Hall and Jones 2007)." Pgs. 5-6, León and Miguel 2016.
    • "Based on household reports on the trade-offs they face between money and walking time to collect water, estimated mean annual valuation for spring protection is US$2.96 per household. Under some stronger assumptions this translates to an upper bound of $0.89 on households’ mean willingness to pay to avert one child diarrhea episode, and $769 on the mean value of averting one statistical child death, or $23.68 to avert the loss of one disability adjusted life year (DALY).
      We believe that the evidence in this article can be interpreted as indicative of relatively low willingness to pay for preventive health among the poor in less developed countries, consistent with other recent work, whereas the precise valuation estimates should be viewed as somewhat more speculative. The link between spring protection and child death—a relatively rare occurrence with multiple possible causes—may be quite difficult for households to discern in practice. The valuation of child health may also differ systematically from adult health (see Davis 2004; Deaton, Fortson, and Tortora 2009).
      Stated preference methodologies, such as contingent valuation, are widely used but controversial (see Diamond and Hausman 1994; Carson et al. 1996; Whitehead 2006; Whittington 2010). We find that although stated preference methods also suggest fairly low valuation, they exceed the revealed preference valuation (which exploits experimental variation in water source characteristics) by a factor of two." Pg. 148, Kremer et al. 2011
    • "Even if we take the largest negative coefficient, which is for the loss of a family member to tuberculosis, the estimate in Table 8 is –0.129, compared with 0.374 for log income, so that the income equivalent of losing an immediate family member to TB is a change in log income of –0.129 divided by 0.374, or –0.345, equivalent to a 29 percent reduction in income. Alternatively, the compensation for the loss would be 41 percent of income for as many years as the effect lasts. By the same token, the compensation for a loss to HIV/AIDS or to childbirth (at least in Table 8) is negative. These numbers seem absurd on their own terms, even before we consider comparing those monetary values to similar monetary values from rich countries. And they are almost certainly gross overestimates, given attenuation bias in the estimates of income through measurement error." Pg. 18, Deaton et al. 2009.
    • "If we were to choose to express the effects on sadness and depression in terms of the effects of income, the results would be much larger than for the ladder of life. For example, if we were to take –0.04 as a representative estimate of the effects of disease, and 0.025 as a representative estimate of the effect of log income, the ratio is –1.6, so that the effects of the death on sadness or depression would be reproduced by an 80 percent reduction in income, or offset by a fivefold increase. Of course, if we were to decide to use these numbers to calculate the monetary equivalent of the death of a family member, we would have to explain why they are to be preferred over the much smaller (and barely significantly different from zero) numbers that come from looking at the ladder of life, particularly given that the ladder is much closer to the life satisfaction measures that have been used in the previous literature." Pg. 19, Deaton et al. 2009.
  • 15.

    For examples of two of the methodologies behind these LMIC-context estimates, see Kremer et al. 2011 and Deaton et al. 2009:

    • Kremer et al. 2011 uses estimates of people's willingness to pay for a water spring protection program to infer the value of life. Its analysis relies on the assumption that households are aware of the preventive health benefits of water spring protection for reducing diarrhea mortality.
      • "Second, we provide among the first revealed preference estimates of the value of child health gains and a statistical life in a poor country. Our estimates fall far below those typically used by public health planners in assessing cost effectiveness and suggest that the demand for health is highly income elastic, as argued by Hall and Jones (2007)." Pg. 146, Kremer et al. 2011.
      • "Under the assumption that households are aware of the relationship between spring protection and diarrhea, combining the results from Tables IV and VI yields an upper bound on the willingness to pay to avert child diarrhea. The bound will be tight to the extent that households’ valuation of spring protection is entirely due to its impact on real and perceived child health, rather than also being due to other spring protection amenities (water clarity, ease of collection, or health gains other than child diarrhea); if these other factors are important, actual willingness to pay to avert child diarrhea will be lower than our estimates. Note that to the extent that people in comparison springs switch to treatment springs in response to the program, we will underestimate both the impact on health and the valuation of spring pro- tection, but to a first-order approximation, both underestimates would be of the same magnitude, so we would not necessarily underestimate health valuations.
        If households have difficulty identifying the links between spring protection and diarrhea, or diarrhea and mortality, we may not correctly estimate the valuation of child health using this approach. In a context in which there are multiple environmental channels for the transmission of fecal–oral diseases (e.g., low rates of handwashing and open defecation, particularly by children), it is plausible that the benefits of an improvement along only one dimension are difficult for households to assess. A 20% reduction in diarrhea prevalence, while biomedically important, might imply a change from five diarrhea cases per year to four cases for a typical child, which would be difficult for a parent to detect. Similarly, the cause of death is difficult to link to diarrhea alone; indeed, for children there is often endogenous feedback between diarrhea and malnutrition.
        Spring protection averts an average of (0.047 diarrhea cases/ child-week) * (1.3 children age 3 and under/household) * (52 weeks/ year) = 3.2 diarrhea cases per household-year. Using our mean spring protection household valuation of 32.4 work days (from the mixed logit), this corresponds to a willingness to pay of 10.1 work days per case of child diarrhea averted. Under the further assumption that spring protection reduces diarrhea mortality by the same proportion as diarrhea incidence, this yields an upper bound on the valuation of a statistical life of 8,742 work days or 35 work years (at 250 work days per year). This bound will again be tight if households’ valuation of diarrhea reduction is entirely due to its impact of mortality.20
        Using the household time values derived from our surveys, and returning to the ITT health impacts estimated in Table IV, the upper bound on the value of averting one case of child diarrhea is a mere US$0.89 (= $0.088 * 10.1 working days), and on avoiding a child diarrhea death is $769 (= $0.088 * 8,742 working days). Using Monte Carlo methods, we estimate a 95% confidence interval ranging from $555 to $1,281. Using the same parameter values to convert diarrhea cases to DALYs as used in the calculations of the value of a statistical life and disability weights proposed by Lopez et al. (2006), the $769 figure corresponds to an upper bound on the value of averting one DALY of about $23.68. Using the higher time value (25% of the average western Kenyan wage) translates into $2,715 per averted child diarrhea death and $83.61 per DALY. These latter figures are likely to be upper bounds on true valuations because water is collected by women and young children who are likely to have much lower than average wages.
        This revealed preference bound on the willingness to pay per DALY averted is far below the cost-effectiveness cutoffs usually used in analyses of health projects in less developed countries. For example, the 1993 World Development Report identified health interventions that cost less than $150 per DALY as “extremely cost effective” (World Bank 1993), and others have used a threshold of $100 per DALY (Shillcutt et al. 2007). Sachs (2002) has argued for setting health cost effectiveness thresholds per DALY at levels corresponding to countries’ gross domestic product per capita, which for Kenya would be over $400, nearly 20 times higher than our preferred estimate. Although an important source of uncertainty in our valuations is the conversion from the value of time to monetary value, it is worth noting that even if our preferred time values were quadrupled, the implied valuation of health and life would still fall below those typically used by public health planners.
        These value of life estimates are also far below the estimated value of a statistical life in the United States and other rich countries (using hedonic labor market approaches), where values typically range from $2 to $7 million (Viscusi and Aldy 2003). Studies from two poorer countries (India and Taiwan) yield estimates on the order of $0.5–1 million, although they are difficult to compare to our sample because they rely on data for urban factory workers. Deaton, Fortson, and Tortora (2009) also find low values of life in African samples using a subjective life evaluation approach. We are unaware of hedonic value of statistical life estimates from the poorest less developed countries.
        Our revealed preference estimate of the value of health is consistent with models in which there is a high income elasticity of demand for health, and thus where households’ valuation on life in less developed countries is very low. Hall and Jones (2007) use US$3 million to $6 million as benchmarks for the value of life in the United States. In a calibration of their model (using data from UNDP 2007), in which the value of a year of life is roughly proportional to per capita annual consumption raised to the constant relative risk aversion utility function curvature parameter (which Hall and Jones suggest plausibly takes on a value of two), the value of a statistical life in Kenya ranges from $953 to $2,711. If per capita consumption in our rural study site is only four fifths of the Kenyan national average, this range becomes $477 to $1,603, accommodating our revealed preference estimate of $769.
        Establishing the ideal way to conduct welfare analysis here is important but beyond the scope of this article, and thus we present a variety of approaches in Section V. We first present results following the conventional “neoclassical” approach of valuing lives according to households’ own revealed preference measures. We then consider the case of a social planner with a $125/DALY valuation (whom we call “paternalistic,” for convenience). This may be appropriate, for example, if the planner values averting child diarrhea more than other forms of household consumption, if children receive less weight in the household welfare function than in the planner’s welfare function, or if households consider only private benefits of reducing diarrhea and ignore disease externalities.
        Using higher spring protection valuations might also be appropriate if households systematically underestimate the health benefits or if they are subject to time inconsistency problems." Pgs. 187-190, Kremer et al. 2011.
    • Deaton et al. 2009 analyzes the reported well-being of individuals whose family members recently died in order to compare the negative impact of this event with the estimated positive impact of additional consumption on reported well-being.
      • "In this section, we take a more direct approach by looking at the effects on respondents of knowing someone who has died. In the 2006 sub-Saharan Africa module of the Gallup World Poll, respondents were asked “Do you personally know anyone that has died from X?” where X includes tuberculosis, malaria, HIV/AIDS, smallpox, polio, hepatitis, and cholera. In the 2007 round, with an overlapping group of countries, the question was changed to “Please tell me if any one in your immediate family has died from X in the past 12 months?” where X includes the same diseases as before plus death from chronic (more than six months) diarrhea and deaths of women in childbirth. In countries where people often have little contact with doctors or clinics, some of these diagnoses are manifestly unreliable, but at the least they provide an indication of how people perceive the effects of these diseases. In interpreting the usefulness of these answers, it should also be kept in mind that reliable data on adult mortality are almost completely absent in many of the countries covered here, and even official estimates of mortality from HIV/AIDS are little more than intelligent guesses based on small surveillance sites or projections from infection rates from surveys or ante-natal clinics." Pg. 10, Deaton et al. 2009.
      • "Table 6 presents the first evidence on the effects of having recently lost an immediate family member on five measures of self-reported well-being. The measure of loss we have used is whether the respondent has lost an immediate family member in the last twelve months to one of (a) malaria, (b) TB, (c) HIV/AIDS, and (d) death in childbirth. The World Poll does not have a question on all-cause mortality, and we have ignored the other reported causes (hepatitis, cholera, polio, smallpox, and chronic diarrhea) because the fractions reporting are very small, and because there are substantial numbers of missing values from “don’t knows” or refusals. We look at five different measures of wellbeing (described in Section 2): (a) the Cantril ladder of life, (b) enjoyment, (c) smiling, (d) sadness, and (e) depression. Each of these measures captures a potentially different aspect of feelings and of life assessment and we have no prior expectation that they will respond in the same way to the deaths of family members.
        Table 6 shows the differences in the SWB measures between people who report having lost a family member and those who do not." Pg. 13, Deaton et al. 2009.
      • "Even if we take the largest negative coefficient, which is for the loss of a family member to tuberculosis, the estimate in Table 8 is –0.129, compared with 0.374 for log income, so that the income equivalent of losing an immediate family member to TB is a change in log income of –0.129 divided by 0.374, or –0.345, equivalent to a 29 percent reduction in income. Alternatively, the compensation for the loss would be 41 percent of income for as many years as the effect lasts. By the same token, the compensation for a loss to HIV/AIDS or to childbirth (at least in Table 8) is negative. These numbers seem absurd on their own terms, even before we consider comparing those monetary values to similar monetary values from rich countries. And they are almost certainly gross overestimates, given attenuation bias in the estimates of income through measurement error." Pg. 18, Deaton et al. 2009.
      • "If we were to choose to express the effects on sadness and depression in terms of the effects of income, the results would be much larger than for the ladder of life. For example, if we were to take –0.04 as a representative estimate of the effects of disease, and 0.025 as a representative estimate of the effect of log income, the ratio is –1.6, so that the effects of the death on sadness or depression would be reproduced by an 80 percent reduction in income, or offset by a fivefold increase. Of course, if we were to decide to use these numbers to calculate the monetary equivalent of the death of a family member, we would have to explain why they are to be preferred over the much smaller (and barely significantly different from zero) numbers that come from looking at the ladder of life, particularly given that the ladder is much closer to the life satisfaction measures that have been used in the previous literature." Pg. 19, Deaton et al. 2009.
      • "Consider first our findings on the value of life in sub-Saharan Africa, and suppose for the moment that it is appropriate to use the life evaluation measures in this way, an issue to which we will return. Given this, we find very small numbers. The largest estimates are 30 to 40 percent of income, and even those estimates are biased upwards by errors of measurement in income. These compensations refer to annual income for the death of an immediate family member in the past twelve months; we have no information on the required compensation in subsequent years. In a comparable exercise for Britain, using data from 1992 to 2002, Oswald and Powdthavee (2009) estimate compensation for the loss of a family member to be between ₤200,000 (upper end estimate for loss of a partner) and ₤16,000 (lower end estimate for loss of a sibling) with monetary amounts in 1996 prices. Median earnings in 1997 were approximately ₤12,500. Viscusi and Aldy (2003) review estimates of the value of a statistical life; these are based on the now standard methodology, dating back to Rosen’s (1988) formulation, in which a value of life is inferred from the earnings premium that workers receive in riskier jobs. For the US, their central estimate for the value of a statistical life is $6.8 million for a prime age worker earning $26,000 a year, or more than 250 times annual earnings. They review comparably-based estimates from around the world—though none from Africa—and estimate that the international income elasticity of the value of life is 0.6 to 0.8, which would imply that the ratio of the value of life to income will be higher in lower income countries. The theoretical concept underlying these estimates is the value of a person’s own life, which is arguably higher than the value of the life of a family member, but they nevertheless provide an indication of the magnitude that is used in the literature and by various government agencies." Pgs. 21-22, Deaton et al. 2009.
      • "The immediate issue is that we have two different measures of wellbeing, a life evaluation measure for which the monetary compensation for a death is low, and affect measures, for which the monetary compensation for a death is large. The ladder question requests an overall evaluation of life; this, or the related question about life satisfaction, is often loosely referred to as “happiness” and has a more plausible claim than momentary feelings or emotions to be a comprehensive measure of individual wellbeing. Yet the affect measures yield more plausible measures of compensation. If we are to decide between them, or possibly rule both to be incorrect, we need a better understanding of what these measures tell us.
        The ladder is an evaluation of life as a whole, affected by momentary experiences and feelings, but distinct from them, Kahneman and Riis (2005). One interpretation is that the ladder is a measure of life achievement, in which material success, education, and social standing are the key ingredients. If so, it is easy to imagine why someone who has lost a parent, for example, could be sad and depressed, but would not necessarily downgrade his or her sense of achievement in life, though we would hardly expect this to be true for the death of a partner or of a child. It is even possible that a sense of dealing well with the misfortune might lead to an improvement in life evaluation; it is possible to wake up in the morning feeling depressed, but still believe that one’s life as a whole is going well (Annas, 2004). If this argument is accepted, neither life evaluation nor life satisfaction measures, informative though they may be, are useful for calculating the compensation for emotional distress. To quote Annas, “if you rush to look for empirical measures of an unanalyzed ‘subjective’ phenomenon, the result will be confusion and banality.” Here the “banality” is our finding that the loss of an immediate family member makes people sad, while the “confusion” is that this sort of unhappiness is the same thing as “happiness” measured as life evaluation.
        A reasonable position is one in which both life evaluation and affect are both components of wellbeing, without having an exclusive claim either separately or together; it is good to have a sense of achievement, and it is good not to be depressed, but other things—such as health—matter too, even if they are not fully captured in either a sense of achievement or in a lack of depression. Any argument for focusing on either affect or life evaluation would also need to deal with the imperfections of each. Affect measures are subject to adaptation, and are easily influenced by trivial features of the situation, while life evaluations often misremember the affective content of past episodes (Kahneman, Wakker, and Sarin, 1997). We are surely on safer ground if we take a capability approach, through which we value health, or income, or other things, by the opportunities for freedom that they provide (Sen, 2001). Improving health extends capabilities, even if those capabilities are not adequately captured by self-reported wellbeing." Pgs. 25-27, Deaton et al. 2009.
      • Deaton et al. 2009 also noted that surveys conducted in African countries that ask respondents to rank policy outcomes often find that people place a higher priority on increasing income than improving health: "Before addressing that question, it is worth considering another measure, previously reported by Tortora (2008). The World Poll asked respondents in Africa to rank in importance twelve objectives based on the Millennium Development Goals. The objectives were (1) providing more jobs for youth, (2) achieving primary education for all, (3) reducing the spread of malaria and TB, (4) improving access to safe drinking water, (5) reducing the death rate among children under five, (6) reducing poverty, (7) reducing the number of women dying during childbirth, (8) reducing the spread of HIV/AIDS, (9) achieving gender equality and empowering women, (10) improving access to sanitation facilities, (11) providing access to new technology, and (12) reducing hunger. Each respondent was given a random selection of six of the twelve objectives, and asked to rank them from one (most important) to six (least important).
        Tortora’s Table 1 shows that reducing poverty and reducing hunger handily win this race, with average ranks of 2.41 and 2.48 respectively. Next, but with a considerably lower rank, comes reducing the spread of HIV/AIDS, with an average rank of 3.05, followed by jobs for youth (3.17), reducing the death rate from children under five (3.34), reducing deaths in childbirth (3.38), achieving primary education for all (3.62), reducing the spread of malaria and TB (3.64), and improving access to safe drinking water (3.75). There is then another substantial gap in the rankings before we come to improved sanitation (4.09), gender equality (4.38), and providing access to new technology (4.65).
        Kharas (2008) reports similar findings from the Afrobarometer surveys from Kenya, Mozambique, Nigeria, South Africa, Tanzania, Uganda and Zambia, where respondents listed their top priorities as jobs, income, support for agriculture, and improvement in infrastructure, with health, including HIV/AIDS, attracting much lower rankings." Pg. 20, Deaton et al. 2009.
  • 16.

    See WHO 2013, "WHO methods and data sources for global burden of disease estimates 2000-2011"

  • 17.

    "The specific measure of cost-benefit analysis adopted in this volume is cost-effectiveness. Effectiveness is measured in natural units (deaths averted and years of life saved) and in disability-adjusted life years (DALYs), a composite measure that combines years lived with disability and years lost to premature death in a single metric (see chapter 15 for an explanation of how DALYs are calculated)." Pg. 36, Disease Control Priorities in Developing Countries, "Chapter 2: Intervention Cost-Effectiveness: Overview of Main Messages"

  • 18.

    For an illustration of how much of a difference age-weighting and time discounting could make to the estimated value of averting a death, see Table 2.1, Pg. 6, WHO 2013, "WHO methods and data sources for global burden of disease estimates 2000-2011"

  • 19.
    • "3% discounting and non-uniform age weighting was used in the original GBD 1990 study. These adjustments result in less weight given to years lived at young and older ages.
    • The GBD 2001-2 study used 3% discounting but uniform age weighting.
    • The GBD 2004 update used the original 3% discounting and non-uniform age weighting.
    • GBD 2004 estimates of DALYs are available for standard DALYs (3% discounting and age weights), no frills DALYs (no discounting, no age weights) and discounted DALYs 3% discounting, no age weights) here."

    WHO, "Disability weights, discounting and age weighting of DALYs"

  • 20.
    • "In summary, we have chosen to simplify the calculation of disability-adjusted life years (DALYs). First, we developed a new normative standard life table for males and females to compute YLLs at each age by identifying the lowest observed death rate for any age group in countries of more than 5 million in population. The new reference life table has a life expectancy at birth of 86·0 years for males and females. Second, years lived with disability (YLDs) have been estimated taking into account comorbidity in individuals. Third, we have computed YLDs simply as the prevalence of each sequela multiplied by the relevant disability weight adjusted for comorbidity. Fourth, on the basis of many arguments,24–26 we have chosen not to discount YLLs, YLDs, or DALYs for time. Fifth, we conclude that we should treat a year of healthy life as equal irrespective of the age at which it is lived. The simpler version of YLLs, YLDs, and DALYs is thus conceptually grounded and easier to explain. It does, however, imply a substantial shift towards greater weight being given to deaths at younger ages, especially younger than 5 years, and greater weight to deaths compared to non-fatal health loss." Pg. 2064, Murray et al. 2010, "GBD 2010: design, definitions, and metrics"
    • "A simpler form of DALY, used by the GBD 2010 study (Murray et al, 2012b), has been adopted.
      This form is easier to explain and use (see Section 2). Age-weighting and time discounting are
      dropped, and the YLDs are calculated from prevalence estimates rather than incidence
      estimates. YLDs are also adjusted for independent comorbidity." Pg. 2, WHO 2013, "WHO methods and data sources for global burden of disease estimates 2000-2011"
  • 21.
    • "The original GBD 1990 study and subsequent WHO updates also incorporated age-weighting in the standard DALYs used in most publications and analyses. The standard age weights gave less weight to years of healthy life lost at young ages and older ages (Murray, 1996). With the clearer conceptualization of DALYs as purely a measure of population health loss rather than broader aspects of social welfare, it is difficult to justify the inclusion of age weights, and the GBD 2010 study dropped them (Murray et al, 2012b; Jamison et al, 2006b) has argued for an alternate form of age-weighting, for incorporating stillbirths and deaths around the time of birth into the DALY. This modifies the loss function for years of life lost for a death at a given age (or gestational age) to reflect “acquired life potential”, by which the fetus or infant only gradually acquires the full life potential reflected in the standard loss function. Murray et al (2012c) have argued that such considerations should be reflected in social priorities rather than in the basic health measure itself.

      Following informal consultations in 2012, WHO decided to adopt the same approach as GBD 2010 in computing DALYs with a time discount rate of 0% and no age-weighting. This change results in a substantial increase in the absolute number of DALYs lost (Table 2.3) and a relative increase in the share of DALYs at younger and older ages (Table 2.2 and Figure 2.1)." Pgs. 7-8, WHO 2013, "WHO methods and data sources for global burden of disease estimates 2000-2011"

    • "Professor Schroeder believes that global health scholars typically do not incorporate age-weighting or adjustments for other ethically-important factors into their 4 estimates of, e.g., the global burden of disease, because they do not see it as their role to make ethical judgments. So, he would not interpret the fact that age-weighting is not typically used as an indication that global health scholars believe it should not be used when making resource allocation decisions. Rather, they believe that judgments about the relevance of such factors should be left to decisionmakers.

      It is potentially problematic that researchers usually do not include moral judgments in their work. Because of the complexity of researchers’ models, policymakers may not realize that moral judgments have not been accounted for, and even if policymakers noticed and wanted to factor them in, they may not have the technical knowledge to revise the models to incorporate their ethical values." GiveWell's non-verbatim summary of a conversation with S. Andrew Schroeder, September 13, 2016.

  • 22.
    • "OECD (2004) reviewed the evidence on economic valuation of mortality among children, and concluded that children have neither the cognitive capacities nor financial resources to state reliable preferences in SP surveys. Thus, society’s perspective is the best perspective from a policy point of view, but it is not applied to children’s preferences – due to difficulties in distinguishing between paternalistic3 and non-paternalistic altruism (and thus the problem of double-counting due to altruism). With paternalistic altruism, it would be appropriate to add-up WTP across individuals. Therefore, parents are asked about the value they attribute to their children’s mortality risk. Some studies find the values of children’s health benefits to be higher than those of adults, while others find the two values to be similar, and one study even finds the value to be less. For further information on SP surveys of parents WTP to reduce mortality risks for their children, see e.g. Alberini, Chiabai and Tonin, 2009; Ferrini et al., 2009; and Scasny, Alberini and Chiabai, 2009. Based on existing reviews of the US and European empirical evidence, it is recommended using a higher VSL for children than for adults (see Chapter 6.2)." Pg. 23, OECD, "Mortality Risk Valuation in Environment, Health and Transport Policies."
    • "The reluctance to make age adjustments of VSL in the United States stems from the significant controversy that erupted over the so-called “senior discount”, where the US EPA used a lower VSL for older individuals in sensitivity analyses conducted for air pollution rules prior to 2004, including the Clear Skies Initiative, where benefits to senior citizens constituted the majority of the policy benefits (Robinson, 2007). Because environmental policies often reduce risks to the very young or the very old, the age differentiation with regards to VSL arose first in this sector. Aldy and Viscusi (2007) note that negative direction of the change in valuation of older people’s lives, rather than recognition of heterogeneity in VSL, may have accounted for the public uproar that the benefit assessment created. If the US EPA had instead placed a premium on the lives of children whose risks would be reduced by the policy, it is likely that few would have objected. Aldy and Viscusi op. cit. also point out that whether VSL should vary by age is not a matter of equity or political expediency, but should rather be grounded on estimates of how people’s WTP for risk reductions vary with age. As people age, their life expectancy shortens, but their economic resources vary as well, giving rise to a theoretical indeterminacy in the age-VSL relationship (see also Viscusi, 2009).
      While there is some empirical evidence that VSL declines at older age, recent work suggest this relationship is uncertain (Hammitt, 2007; Aldy and Viscusi, 2007; Krupnick, 2007). Thus, determining the VSL at different ages requires more research. Age differentiation in VSL will facilitate better prioritisation of mortality risk reduction efforts for populations of various ages. Two US expert panels have advised against making VSL age adjustments due to inconclusive evidence (Cropper et al., 2007; National Academy of Sciences, 2008).
      The meta-analysis of SP studies of adult VSL in Chapter 3 found no clear relationship between age and VSL, although for a subset of the data, indications of an inverted U-shape relationship between VSL and mean age of the sample was found (meaning that VSL increase with age to about 40-50 years of age and then decline, see Annex 3.A1).
      VSL appears to be higher for children, due to parents’ altruistic concerns for their children, with results from the United States and Europe indicating VSL for children being as high as a factor of 2 that of their parents/adults (US EPA, 2003; OECD, 2010). More generally, in cases where the policy intervention particularly affects children, due to the nature/scope of policy (e.g. pesticides in school grounds) or because children are particularly vulnerable to this particular hazard (e.g. lead in drinking water), then child-specific values are likely to be particularly helpful in ensuring that resources and policy efforts are allocated efficiently. According to OECD (2010), it is likely that the introduction of a “premium” for children would raise less controversy than a “discount” for seniors. Since “children” were not included in the studies used to determine baseline VSLs, the “premium” could be simply added to the baseline estimate. Moreover, there is a stronger political case. While the interests of children are usually defended by parents (and other care-givers), policy makers in OECD governments have always had a special role in protecting the interests of children with respect to risks in general. In some cases (i.e. negligence or abuse), this role may supersede that of their parents. As such, there is, at least, a distinct obligation with respect to children’s risks to determine whether or not a premium should be applied.
      Based on literature reviews and the SP meta-analysis of Chapter 3, no adjustment for age is recommended. However, when the policy that is analysed targets children specifically (or affects mainly children), a higher VSL for children is recommended, based on the available empirical evidence from the United States and Europe (US EPA, 2003; OECD, 2010). VSL for children should be 1.5-2.0 times higher than the mean adult VSL." Pgs. 130-131, OECD, "Mortality Risk Valuation in Environment, Health and Transport Policies."
  • 23.

    "The reluctance to make age adjustments of VSL in the United States stems from the significant controversy that erupted over the so-called “senior discount”, where the US EPA used a lower VSL for older individuals in sensitivity analyses conducted for air pollution rules prior to 2004, including the Clear Skies Initiative, where benefits to senior citizens constituted the majority of the policy benefits (Robinson, 2007). Because environmental policies often reduce risks to the very young or the very old, the age differentiation with regards to VSL arose first in this sector. Aldy and Viscusi (2007) note that negative direction of the change in valuation of older people’s lives, rather than recognition of heterogeneity in VSL, may have accounted for the public uproar that the benefit assessment created. If the US EPA had instead placed a premium on the lives of children whose risks would be reduced by the policy, it is likely that few would have objected. Aldy and Viscusi op. cit. also point out that whether VSL should vary by age is not a matter of equity or political expediency, but should rather be grounded on estimates of how people’s WTP for risk reductions vary with age. As people age, their life expectancy shortens, but their economic resources vary as well, giving rise to a theoretical indeterminacy in the age-VSL relationship (see also Viscusi, 2009).
    While there is some empirical evidence that VSL declines at older age, recent work suggest this relationship is uncertain (Hammitt, 2007; Aldy and Viscusi, 2007; Krupnick, 2007). Thus, determining the VSL at different ages requires more research. Age differentiation in VSL will facilitate better prioritisation of mortality risk reduction efforts for populations of various ages. Two US expert panels have advised against making VSL age adjustments due to inconclusive evidence (Cropper et al., 2007; National Academy of Sciences, 2008).
    The meta-analysis of SP studies of adult VSL in Chapter 3 found no clear relationship between age and VSL, although for a subset of the data, indications of an inverted U-shape relationship between VSL and mean age of the sample was found (meaning that VSL increase with age to about 40-50 years of age and then decline, see Annex 3.A1)." Pgs. 130-131, OECD, "Mortality Risk Valuation in Environment, Health and Transport Policies."

  • 24.
    • See Figure 1, Pg. 4, Jamison 2016.
    • "In order to assess the relative value of fetal loss and/or death at various ages without directly triggering an attempt by respondents to give the ‘right’ answer, the survey was designed primarily for indirect comparisons, similar to the methodology of [18]. In particular, in each of the ten conditions to which participants were randomly assigned, they were asked a single additional question beyond the six demographic questions included in Table 1. For the eight primary conditions, this took the following form:

      Appropriate life-saving programs can prevent many causes of death. Suppose that there are two different life-saving programs and that they target different age groups of the population.

      Program A saves 100 adult women.
      Program B saves [X] infants in the first week after birth.

      How many infants (can be less than, equal to, or greater than 100) would Program B have to save in order for you to value both programs qually, in the sense of benefiting society as a whole by the same amount? Assume that both programs have the same cost, and please think carefully about your answer.

      Respondents could choose either to specify an exact numeric value X or to state that no number X would be sufficient to make the programs equivalent in their mind. This latter option was included both to allow for the possibility that the value X might be essentially infinite for some people in some cases, and to allow for the possibility that some people would simply be uncomfortable making such comparisons about deaths and lives saved." Pg. 3, Jamison 2016.

    • "The second broad conclusion concerns the actual numeric trends observed in the data. We find that 10-week fetuses are valued less than 39-week fetuses (especially when taking into account the potential selection bias regarding which subjects gave explicit responses), but they are still valued at roughly half the level of adult women. Furthermore, there is no difference in the average valuation of pregnant women at the corresponding gestational stages, although both are higher than for non-pregnant women. This suggests overall that society places a distinctly positive value on fetuses, even quite early in pregnancy, but also that the imputed value increases over time, contradicting a purely binary perspective.
      One-year-old children are valued almost (but not quite) as highly as adult women, especially by male respondents. However, one-week-old infants are valued much less. Indeed, we do not find a difference in valuation between these neonatal infants and full-term fetuses. Biologically this makes sense, but it is edifying that the birth event does not have a stronger impact on valuations. This is true even when we ask subjects to compare the two groups directly, which is a more stringent test of the hypothesis. Overall, then, the perceived value of life is weakly positive and increasing with gestational age; it does not exhibit a discontinuity at birth; around birth it is still substantially below that of an adult woman; but by age one it has narrowed most of the gap." Pgs. 5-6, Jamison 2016.
  • 25.
    • "The proportion of respondents who refused to input an explicit number X (which would make program B equally valuable in their mind to saving 100 women in program A) ranged from 37% to 45% across all treatments, except for the two treatments involving 10-week fetuses or 10-week gestation babies, where it ranged from 56% to 60%. This is suggestive of a situation in which roughly 40-45% of individuals simply do not like to make such comparisons no matter the ages involved, and in which a further roughly 15% of individuals may be willing to make such comparisons in general but believe that an adult woman’s life is worth an arbitrarily large number of 10-week fetal losses. However, this interpretation is not conclusive as it is impossible to verifiably ascertain what led to the various refusals, and throughout the rest of the paper the focus remains solely on those who were willing to input an explicit number X." Pg. 3, Jamison 2016.
    • "To be perfectly clear, there is no indication that the specific values derived from the experiment in this paper should be imported wholesale into policy analyses. However, the qualitative comparisons at the end of the previous paragraph provide strong support for one family of conceptual models that has been advanced in the literature, and against alternative models that violate one or more of the given properties. The magnitudes themselves may also be useful as one input among many when calibrating such models for cost-effectiveness and resource prioritization. Further more refined experiments on a targeted (but potentially broader) sample population, or subgroups thereof such as health professionals, could yield more robust empirical estimates regarding the valuation of life not just around the time of birth but at older ages and across geographic locales." Pg. 6, Jamison 2016.
  • 26.
    • "Philosophers writing on the value of life can roughly be categorized into a few major schools of philosophical thought, including utilitarianism, contractualism, and Kantianism. These viewpoints could yield implicit judgments for how to weigh different health outcomes for different populations, but there is no centralized discussion of this topic in the field. Philosophers do not typically discuss explicit quantitative moral tradeoffs in their work." GiveWell's non-verbatim summary of a conversation with S. Andrew Schroeder, September 13, 2016.
    • "Global burden of disease estimates made by the Institute for Health Metrics and Evaluation and the World Health Organization typically assign the most weight to averting under-5-year-old deaths because this is expected to avert the largest number of years of life lost. This assumption is also dominant in the cost-effectiveness literature generally, and there is pushback to attempts to introduce other assumptions...To date, the discussion of age weighting and other ethical considerations in cost-effectiveness analysis has been limited." GiveWell's non-verbatim summary of a conversation with Dean Jamison, January 9, 2017.
  • 27.
    • For underlying calculations, see this spreadsheet, Sheet "'Standard' vs. actual results comparison". We have excluded the estimate for Sightsavers since it is primarily based on the estimate for the Schistosomiasis Control Initiative.
  • 28.

    See the "Moral weights" sheet for details.

  • 29.
    • "The Global Health 2035 Commission on Investing in Health, co-chaired by Dr. Lawrence Summers and Dr. Jamison, performed a benefits-cost analysis (BCA) and included age weighting of health outcomes in its published report. The commission assigned different weights to stillbirths averted and adult deaths averted, relative to infant deaths averted. These weights are similar to those in chapter six of GBD 2006." GiveWell's non-verbatim summary of a conversation with Dean Jamison, January 9, 2017.
  • 30.

    Malaria Consortium's seasonal malaria chemoprevention program only treats children under 5 years old whereas the Against Malaria Foundation distributes nets to people of all ages, so a larger proportion of our estimate of Malaria Consortium's benefits comes from averting infant deaths. We estimate that the Against Malaria Foundation averts some adult deaths.

  • 31.
    • GiveWell staff sometimes value averting the death of an adult about 1-4x more than the death of an infant, whereas "standard" approaches typically value averting the death of an infant about 1-2x more than the death of an adult.
    • Note that there is substantial variability in moral weight assumptions among individual GiveWell staff members; some staff members may have very different overall estimates under "standard" assumptions. See here for more detail.
    • For more, see the "Moral weights" sheet in our actual model relative to the version of our model that uses only relatively "standard" moral weights.
  • 32.
    • The cost-effectiveness of deworming relative to cash transfers was barely affected (deworming was ~7.9x cash with "standard" moral weights and ~8.1x cash with actual staff moral weights) because the primary impact of both cash transfers and deworming programs is on income, so the relative moral value of benefits from these programs was unchanged.
    • The bottom line cost-effectiveness for the Against Malaria Foundation (AMF) did not change considerably due to offsetting effects. Our AMF analysis includes its impacts on under-5-year-old deaths, adult deaths, and income (via improved childhood development). Relative to GiveWell staff, "standard" assumptions generally assign substantially higher weight to under-5-year-old deaths, substantially less weight to adult deaths, and slightly less weight to income relative to health. But, these differences roughly offset each other.
      • Under "standard" moral weight assumptions, averting the death of an under-5-year-old is about 91x as valuable as doubling consumption for one person for one year. (For calculation, see Cell B46, Sheet "Moral weights.") In the May 2017 version of our model, most staff inputs implied that averting the death of an under-5 year old was about 10-65x as valuable as doubling consumption for one person for one year. (See Row 36, Sheet "Moral weights", which draws from Row 29, Sheet "Moral weights" on the published version of our model.) So, for most staff, the relative value of averting deaths for under-5-year-olds became ~1-9x more valuable.
      • However, for most staff, much of the potential benefit of AMF came from increasing income and averting deaths of over-5 year olds, both of which became relatively less valuable under "standard" assumptions.