How we measured the value of a statistical life in Kenya and Ghana

Why measure people’s preferences?

Many of us intuit that human lives are sacred and priceless. It feels difficult and even repugnant to put a dollar value on someone’s life, whether our own, a loved one’s, or a fellow citizen’s. However, sometimes this calculation is necessary for effective policymaking: with limited resources, governments have to decide whether to avert deaths (for example by setting speed limits, putting in place safety standards or pollution regulations, etc.) or prioritize other outcomes like economic output.

This also applies to decision-makers in international development who aim to maximize the impact of aid dollars on human wellbeing. For instance, GiveWell needs to decide how to allocate donor money among different programs, including bed nets that save lives and cash transfers that can increase a family’s food consumption and that of other goods. To improve wellbeing, it is important to account for the preferences of people on the receiving end of the program, including how they value reducing mortality risks versus increasing income. However, as this blog explores, there are very few studies illustrating the preferences of people in poverty, which motivated our study in partnership with GiveWell.

Should we measure people’s program (cash transfer) or outcome (increasing income) preferences?

The first challenge is whether to measure which programs or which outcomes people prefer. There are studies on how people trade-off between cash transfers versus bed nets (like this one), and studies on how people trade-off between reducing mortality risks and increasing income (like this one).

In principle, how people value programs (e.g. cash transfers vs. bed nets) is driven by the outcomes they value (e.g. income vs. mortality risks), and we could deduct one from the other. In practice, they may diverge for the following reasons:

• Information: People may not fully understand how a program affects various outcomes. For instance, when forming valuation of bed nets, they may not know how bed nets reduce malaria mortality. Similarly, it may be difficult to predict the average impact of cash transfers on a person’s welfare — even as a potential recipient.
• Behavioral biases: People spend much more on medical treatments than on preventions, which may suggest inconsistent preferences, possibly driven by people’s tendency to value the present more than the future¹.
• Other aspects of the program: Even if people had perfect information and consistent preferences, their preferences for a program could be driven by factors other than their valuation of averting mortality and morbidity, e.g. ease or comfort using a bed net, taste of water treated by chlorine dispenser. While these are legitimate factors to consider when analyzing a person’s perspective of a service, they pose a challenge when trying to disentangle what someone values from their feelings about a program².

For these reasons, we decided to focus on people’s preferences of outcomes, which will also allow us to derive valuation for a number of different programs achieving similar outcomes (e.g. all health programs that avert mortality and morbidity).

In addition to studies eliciting valuations of programs or products (e.g. this, this and this), there are a number of studies measuring how people rank policy goals or programs (e.g. this and this). While interesting, their results are driven by complex factors beyond people’s preferences over outcomes,³ and they don’t necessarily produce enough data on preferences to be directly applicable for a decision-maker looking to quantitatively decide between programs. For example, knowing that Sub-Saharan Africans rank “reducing poverty” above “reducing the death rate among children under five” does not tell us whether a government should increase or decrease its spending on anti-poverty programs relative to child health programs — it depends on whether they’re spending too much or too little compared to what people desire, how much it costs to reduce poverty or death rates by a certain amount, and other factors. Therefore, we see value in collecting data on how people trade-off among outcomes.

Why not just give cash?

Some argue that if donors in international development want to maximize recipient wellbeing by respecting their preferences, they should just give cash.

Cash transfers are a great service that is relatively easy to implement using mobile money in many places and has demonstrated positive impact. It allows people to make decisions based on their preferences as well as knowledge about local market conditions and their own skills. However, it is possible that even the best way to respect people’s preferences involves more than just giving cash. For example, even if people place a high value on averting mortality, they may not end up buying health products that are effective in averting mortality (e.g. bed nets or Vitamin A supplements) once they are given cash. This could be because of low demand for preventative health products, as explained above, or because these products are not available in the local market without non-profits distributing them. In addition, people may undervalue health products that have positive externalities. In this case, it is worthwhile distributing health products (potentially free or subsidized, if households heavily undervalue them) in addition to giving cash.

Our methodology for measuring the value of statistical life

Policymakers are usually concerned about valuing mortality risks since most programs do not save a life or cause death with full certainty. People’s valuation of small changes in mortality risks is denoted by the concept of value of statistical life (VSL). VSL is used by government agencies including the US Environmental Protection Agency (US EPA) and the US Department of Transportation to guide policy decisions. (In an upcoming blog post, we will discuss another approach we used in the study, which focuses on eliciting people’s “moral preferences” over outcomes for the community rather than “private preferences” over outcomes for themselves and their household).

There are two approaches to estimating VSL: the revealed preference approach — using real choice data and the stated preference approach — using hypothetical questions. The US EPA aggregates results from both types of studies to arrive at their VSL estimate, while some other OECD countries rely more on stated preference methods.

Revealed preference

This approach uses real-life choices made by people facing different options with trade-offs between money and mortality risks, for example, professions with mortality risks and different wage levels (hedonic wage), vehicles with different prices and safety features and whether to purchase an airbag (consumer choices).

Advantages of this approach

The main advantage of this approach is that it is from real choices and hence reflects people’s true preferences.

Challenges with this approach

A major issue with this approach is that people in the sample from which VSL is estimated may not have the correct belief of how much different choices increase or decrease their mortality. Since estimation models assume they have correct beliefs, VSL estimates would be biased if these beliefs are incorrect, and we often have no way to find out⁴. Some of the few exceptions, which measure both beliefs on mortality risks and VSL, are this study from Nepal and this study from Taiwan, and the latter finds evidence for incorrect beliefs on mortality risks⁵.

Another issue is liquidity constraint for willingness-to-pay (WTP) studies⁶. People may exhibit higher VSL if they had the ability to borrow money compared to if they did not, and in practice may not have perfect ability to borrow (note that we are holding lifetime income constant here and only talking about the effect of the ability to borrow). In practice this isn’t an issue for hedonic wage studies, the most common type of revealed preference VSL studies, since they estimate willingness-to-accept (WTA) ⁷.

A potential challenge with the WTA approach is loss aversion. People may need to be compensated more to increase their mortality risk by some amount from their baseline level, compared to their WTP to reduce risk by the same amount from the baseline. This divergence between WTP and WTA likely reflects loss aversion rather than inherent preference over changes in mortality risks⁸.

Furthermore, people may have different valuations for mortality risks of different nature, and it may be unclear how policymakers should extrapolate across types of risks⁹.

Stated preference

This approach usually involves asking respondents hypothetical questions about WTP for hypothetical products associated with a certain amount of mortality risk reduction¹⁰. It has been widely used to derive values on health and environmental outcomes, including by the US EPA.

Advantages of this approach

• One can focus on the outcome of interest (e.g. mortality risk reduction) rather than other aspects of a health product (e.g. comfort, taste).
• One can give very clear information on the magnitude of the risk reduction, reducing bias in the VSL estimate due to incorrect beliefs.
• One can use the results to assign valuation to public goods¹¹.
• One can study outcomes for which it’s difficult to find an associated product with real choices, e.g. the preservation of nature.

Challenges with this approach

• Hypothetical bias: Since real choices aren’t involved, people may take the choice less seriously and give a value that would be too high or low compared to what they would actually pay.
• Other biases associated with hypothetical questions: One example is social desirability bias, where respondents would say something that they think others will see as “good” rather than their true value¹².
• Imperfect comprehension of probabilities: Even if people were given information on the mortality risks, they may not understand it well¹³. We’ll be posting a blog going further into this month.
• Sensitivity to the type of risk: Just as in the revealed preference case, people may have different values of different types of mortality risk, e.g. traffic accidents versus diseases.
• Sensitivity to framing: As we ask about the value of a hypothetical product, the description of the product (e.g. medicine vs. vaccine) may impact valuation without any change in the nature of mortality risk or other parameters.

For the WTP approach, liquidity constraint is again a challenge: even in hypothetical choices, people may perceive themselves as having imperfect ability to borrow.

Loss aversion is again a potential challenge with the WTA approach.

What we knew and our prior

VSL from high and low-middle income countries

There are many studies on VSL in high-income countries (HICs). The US EPA aggregates across 10 revealed preference and 10 stated preference VSL studies from the US that satisfy their criteria and arrives at the value of 9.7 million USD for the US¹⁴. (These are obtained by estimating WTP to reduce mortality risk by a small amount and then scaling by dividing by the risk reduction level. However, one should not think of this as the value of saving a life with 100% probability).

There have been numerous studies on VSL using revealed and stated preference approaches, including some from Low- and Middle-Income Countries (LMICs). This review identifies 8 revealed preference and 17 stated preference studies in LMICs. Very few VSL studies have been conducted in Sub-Saharan Africa. There are two revealed preference studies: one on people travelling in different means of transportations to an airport in Sierra Leone, and one on people’s willingness to walk to get clean water in Kenya. In addition, there are two stated preference studies: one from Tanzania and one from Sudan.

Most of the above studies do not target people living in extreme poverty, a very relevant population for decision-making in international aid¹⁵.

Challenges extrapolating VSLs across different contexts

In principle, evidence from other contexts can inform VSL values in our population of interest, namely low-income populations in LMICs. In practice, the consequent predictions span a very wide range of values, calling for further collection of empirical evidence from this population.

One can extrapolate VSL results from one context to another using a “benefit transfer” exercise. This is done mostly using the relationship between VSL and income, or the income elasticity of VSL¹⁶.

Within-country studies (e.g. this and this) have typically found an elasticity of below 1¹⁷, implying that as income increases VSL increases more than proportionally to income. Cross-country studies including HIC and LMICs find elasticity values close to or above 1. Previous research recommends using an elasticity of 1.5 when extrapolating to low-income settings.

The World Health Organization recommends cost-effectiveness benchmarks based on valuing a life year at 1 to 3 times GDP per capita, a roughly linear extrapolation. GiveWell’s recommendation prior to 2019 was consistent with valuing a life year at 2.5 times per capita annual consumption.

To understand why these calculations can be tricky, we offer the below example extrapolating VSL from the US, to a population of typical GiveDirectly recipients in Western Kenya.

Here are a few predicted values assuming:

• US GDP per capita in 2013 (nominal): 53,107 (source)
• US VSL (2013 USD) : 9.7 million (source)

● Annual consumption per capita of Give Directly recipients in Western Kenya (nominal, 2013): 286 (source)

Under different assumptions for benefit transfers recommended here:

1. Elasticity of 1: 52,238 (2013 USD)
2. Elasticity of 1.5: 3,833 (2013 USD)
3. VSL as 100 times annual income: 28,600 (2013 USD)
4. VSL as 160 times annual income: 45,760 (2013 USD)

The benefit transfer exercise yields a wide range of estimates different by an order of magnitude. Moreover, the quantitative trade-off between reducing mortality risk and increasing income has great influence on the relative cost-effectiveness of development programs. This points to the need of collecting data from the relevant populations using methods comparable to those that have been applied elsewhere.

Conclusion

In this post, we have made the case for the importance of measuring people’s trade-offs or preferences over different outcomes we care about (e.g. reducing mortality and increasing income) for policymaking and resource allocation. We also went over two main approaches to capture such preferences ­­– revealed preference, using real choice data, and stated preference, using hypothetical questions — as well as their pros and cons. Finally, we summarized what we knew about VSL prior to our study and made the case for more empirical data collection of VSL in LMICs. In the next post, we will describe what IDinsight did in our study, what we found, and future work we are excited about.