Primary Author: Sanjay Joshi, with thanks to multiple reviewers in SoGive
This was submitted to the GiveWell Change Our Mind Contest.
Exec summary
SoGive believes GiveWell could fruitfully employ more effort researching the discount rate. This document will set out the following points:
This document covers the main GiveWell discount rate of 4% used to discount future economic/financial benefits. GiveWell also uses a discount rate of 0.5% for future averted deaths in the New Incentives model. Because of time constraints, we only focused on the 4% discount rate.
Ideally we would prefer to avoid giving a recommendation for the discount rate which GiveWell should use, since our document highlights several areas where further research is needed. However our non-resilient impressions are outlined below:
Table 1: Highly non-resilient impressions about how GiveWell should change the discount rate
| Consideration | Δ | Comments |
| 3. Discount rate construction method | 0% | We tentatively prefer GiveWell’s method to the Ramsey Formula |
| 4. Pure rate of time preference | 0% | We remain sympathetic to the 0% pure rate, despite our counterarguments. |
| 5. Catastrophic risks | +0.9% | Our estimate of 2.3% was non-resilient, however it’s our best estimate at this time |
| 6. Uncertainty | -0.1% | We expect that the variance in the discount rate is significant enough to warrant a modest decline in the rate |
| Overall | +0.8% | We tentatively recommend a discount rate of 4%+0.8% = 4.8%; this depresses deworming cost effectiveness by c15%. |
Readers may also be interested in our review of the elasticity of the marginal utility of consumption (Appendix 3) and our view that “temporal uncertainty” (i.e. catastrophe risk) is being applied inconsistently between outcomes (Appendix 4).
(1) Small changes in discount rate can cause material changes in cost-effectiveness numbers
GiveWell’s cost-effectiveness model is highly sensitive to the choice of discount rate for deworming charities, and still somewhat sensitive to the discount rate for other charities.
Table 2: Sensitivity analysis showing how cost-effectiveness changes when a 5% discount rate is used instead of a 4% discount rate.
| Impact on Cost Eff as mult of cash* | Impact on Cost eff b4 lev/funging adj** | Impact on Units of value*** | |
| AMF | -4.8% | ||
| GiveDirectly | -2.3% | ||
| Deworm the World | -17.4% | ||
| END fund | -17.6% | ||
| SCI | -18.5% | ||
| Sightsavers | -18.1% | ||
| Malaria Consortium | -3.3% | ||
| Helen Keller | -0.1% | ||
| New Incentives | 1.3% | -0.9% | |
| * "Cost Eff as mult of cash" is short for "Cost effectiveness as a multiple of cash transfers". This is one of the main measures of cost-effectiveness used by GiveWell. | |||
| ** "Cost eff b4 lev/funging adj" = Cost effectiveness before leverage/funging adjustment. It's needed because a quirk of the GiveWell model means the bottom line AMF cost-effectiveness figures can't be compared in this sort of sensitivity analysis | |||
| *** Units of value is used for GiveDirectly because the standard cost-effectiveness ratio uses GiveDirectly's impact as the denominator, so this will automatically be 1x anyway. | |||
In order to calculate this, we took a copy of GiveWell’s model, and adjusted the discount rate in this cell.
(2) Givewell's discount rate is lower than some other actors
GiveWell’s discount rate of 4% is a relatively modest discount rate compared to peers. For example, the World Bank uses discount rates of 10% to 12% and JPAL uses a discount rate of 10% per year. Similarly, a review of the Social Discount Rate for Cost-benefit analysis conducted by the Asian Development Bank observed that developing countries generally apply higher social discount rates (8–15%) than developed countries (3–7%).
It is not necessarily the case that GiveWell should change their discount rate just because peers are different, however it is useful to be aware of this difference. It appears that a material driver of the difference is the non-standard method of constructing the discount rate.
(3) GiveWell uses a non-standard method of constructing a discount rate
GiveWell’s approach to constructing the discount rate was to identify relevant considerations, assign a percentage for each, and then add them up. While having the merit of clarity and simplicity, this is a non-standard approach.
Table 3: How GiveWell constructed their discount rate
| Discount rate | GiveWell |
| Improving circumstances over time | 1.7% |
| Compounding non-monetary benefits | 0.9% |
| Temporal uncertainty | 1.4% |
| Pure time preference (donors) | 0.0% |
| Pure time preference (program participants) | 0.0% |
| Overall discount rate | 4.0% |
[GiveWell expands on the rationale for each component in this document. Further details can also be found in this older (2018) GiveWell document.]
More standard approaches to setting a discount rate include the SRTP (Social Rate of Time Preference) approach and the SOC (Social Opportunity Cost) approach. More information about the SRTP and SOC approaches can be found in Chapter 10 of the book Cost-benefit Analysis by Boardman et al 2018.
An example of the SRTP approach is the Ramsey formula, which is widely used and well known.
Ramsey formula Discount rate = δ + ηg |
δ = pure rate of time preference
η = elasticity of marginal utility of consumption
g = growth rate of real consumption
According to GiveWell’s assumptions:
δ = 1.4%
η = 1.59
g = 3%
So if GiveWell used the Ramsey formula, they would choose a discount rate of
Discount rate = 1.4% + 1.59 ✕ 3% = 6.17% = 6.2% (2sf)
However GiveWell’s actual discount rate is only 4%, which is materially lower than the rate suggested by the Ramsey formula.
(3.1) How GiveWell arrived at 4%, compared to 6.2% from the Ramsey Formula
GiveWell’s discount rate of 4% was calculated as:
Improving circumstances over time + compounding non-monetary benefits + temporal uncertainty
= 1.7% + 0.9% + 1.4% = 4%
The Ramsey Formula, by contrast, uses
δ+g*η=1.4%+3%*1.59=1.4%+4.77%=6.2%
A key difference between these calculations is that the Ramsey formula simply takes g and η and multiplies them together.
GiveWell’s equivalent figure of 1.7% was calculated in this spreadsheet.
The calculations have a clear, intuitive logic to them:
The more common approach (the Ramsey model) would lead to 4.77%:
At first glance, the fact that the Ramsey model assumes that an optimum has been achieved appears to be a weakness, especially in a low income country.
However, on balance, we think that GiveWell’s model does not appear as clear cut as its intuitiveness would suggest.
Because of time constraints we stopped our investigations at this point, however we think that further work on this topic could explore:
If we were to select a view on this at this stage, we would likely side with GiveWell’s approach, given its intuitiveness, but believe that it is unwise to assume that this is correct without further research.
(4) GiveWell appear to have missed reasonable arguments in favour of a pure rate of time preference
We reviewed each of the three components (δ, η, and g), and we think that δ (pure time preference) deserves the most attention out of each of GiveWell’s assumptions.
GiveWell has given relatively little thought to the question of pure time preference
GiveWell allows for zero pure time preference. The rationale is a negative one – objecting to the findings of revealed preference studies. We agree with this objection to beneficiary pure time preference. However we also note that, to the extent that GiveWell donors are within the effective altruism community, GiveWell donors are less likely to be longtermist, which skews the probabilities in favour of them having a positive pure time preference.
All that GiveWell has said on the topic of the pure rate of time preference in their 2020 document is captured in the below box:
|
This is a very small amount of detail, and is in a document which starts with: “Note: This is a rougher stage internal document that has not undergone our formal vetting process.”
The 2018 document has slightly more detail, and includes the sentence “I have not heard a good philosophical argument that we should have a pure time preference, beyond an appeal to intuition.”
Good philosophical arguments in favour of a positive pure time preference do exist
While we are sympathetic to GiveWell’s conclusion, we do not believe there is a total absence of good philosophical arguments in favour of a pure time preference.
The strongest argument in favour is this paper by Andreas Mogensen, based on our currently incomplete review of this topic.
Further work on this topic should explore:
Overall, we are sympathetic to the idea that an appropriate pure rate of time preference may well be zero, however we believe that GiveWell could fruitfully research the question further and may reach a different conclusion.
(5) GiveWell could consider catastrophic risks more thoughtfully
GiveWell’s discount rate includes 1.4% for “temporal uncertainty” which allows for catastrophes. This is based on a logic that if a severely catastrophic event occurs, the intervention’s work would be wasted if the benefits would have arisen after the catastrophe.
GiveWell determined their 1.4% figure relying heavily on subjective judgement
Further research could lead to a materially different number
We think the exercise of determining the 2.3% figure is valuable for illustrating that the 1.4% figure could change non-trivially if more time were spent considering this figure. We are not arguing that we have high confidence in that number.
One more note on the topic of the temporal uncertainty component of the discount rate: arguably the “temporal uncertainty” component of the discount rate is being applied inconsistently between interventions. In order not to interrupt the flow of this document, this argument is expanded upon in Appendix 4.
(6) GiveWell should consider the amount of uncertainty in their discount rates and consider whether declining discount rate considerations apply
Imagine that there is uncertainty about the correct choice of discount rate. It turns out that this should lead to a somewhat lower, and declining discount rate.
Table 4: How uncertainty affects discount rates
| Adjustment to discount rate | Applicability for GiveWell |
| This is often considered to be too modest an effect to be worth modelling over a time period of just a few decades. However if the variance/uncertainty is high enough, this might be worth incorporating. Arguably, given the uncertainties around the temporal uncertainty component (i.e. catastrophe/extinction events), a high variance in the discount rate is justified. |
| This consideration likely won’t be material for GiveWell, since that typically only applies over longer time horizons (e.g. in an intergenerational context). Again, if the variance/uncertainty in the discount rate is large enough, it might be worth incorporating this. |
6(a) Why uncertainty in the discount rate leads to lower discount rates
Imagine a simple model where you don’t know what the right discount rate is, but you know it’s either 3% or 5%.
Table 5: Present Value of a cash-flow of €1000 received after t years

Source: Gollier, Koundouri, Pantelidis (2008)
6(b) Why uncertainty in the discount rate leads to declining discount rates

The forward discount rates
Source: Gollier, Koundouri, Pantelidis (2008)
Arrow et al 2014 states that allowing for uncertainty/risk could lead to a precautionary effect: “that is, that uncertainty about the rate of growth in consumption reduces the discount rate, which causes the social planner to put more weight on the future”. They cite Pindyck and Wang 2013. We reviewed this paper quickly and didn’t find enough detail to justify why the reduction in discount rate makes sense in that context, and whether it carries over to the GiveWell context.
Our tentative view is that there is material uncertainty in the discount rate, especially with regard to catastrophic/extinction risks, and that this likely leads to a modest reduction in the discount rate over the time period relevant for GiveWell.
Useful reference documents
Recent (2020) GiveWell document outlining rationale for current discount rates, by JS
GiveWell calcs to justify the 1.7% discount rate component relating to economic growth, JS
Older (2018) GiveWell document on rationale for discount rates, also by JS
Older (2018) GiveWell document on rationale for discount rates, by CAM
Useful documents about discount rates and the Ramsey Formula:
This 50-page survey from the Asian Development Bank includes useful background on the discount rate
Appendix 1: impact of a 10% discount rate
As noted above, a discount rate of 10%-12% is common. For illustration purposes, we set out here the impact of a 10% on the GiveWell cost-effectiveness model
Table 6: Impact of a 10% discount rate on cost-effectiveness measures
| Cost Eff as mult of cash* | Cost eff b4 lev/funging adj** | Units of value*** | |
| AMF | -14.3% | ||
| GiveDirectly | -11.2% | ||
| Deworm the World | -63.6% | ||
| END fund | -64.7% | ||
| SCI | -67.8% | ||
| Sightsavers | -66.3% | ||
| Malaria Consortium | -8.7% | ||
| Helen Keller | 3.4% | ||
| New Incentives | 9.0% | -3.2% | |
| * "Cost Eff as mult of cash" is short for "Cost effectiveness as a multiple of cash transfers". This is one of the main measures of cost-effectiveness used by GiveWell. | |||
| ** "Cost eff b4 lev/funging adj" = Cost effectiveness before leverage/funging adjustment. It's needed because a quirk of the GiveWell model means the bottom line AMF cost-effectiveness figures can't be compared in this sort of sensitivity analysis | |||
| *** Units of value is used for GiveDirectly because the standard cost-effectiveness ratio uses GiveDirectly's impact as the denominator, so this will automatically be 1x anyway. | |||
For example, the 67.8% decline in the cost effectiveness of the SCI Foundation is because the cost effectiveness multiplier declines from 13.5x cash to 4.3x cash.
Appendix 2: An alternative view on the probability of sufficiently catastrophic risks
This is a quick attempt at constructing an alternative view for what GiveWell calls temporal uncertainty.
The below model tentatively proposes an alternative figure of 2.3%, however as can be seen from the comments in the below table, this was based on a time-constrained review, and hence could be improved with further work.
The table sets out the work done by one member of the SoGive team, and is not necessarily representative of SoGive’s house view.
We believe that the value from the below exercise is not to suggest that 2.3% is, in some sense, the “right answer”. Rather this demonstrates it’s possible to come to conclusions which are materially different from the 1.4% figure employed by GiveWell.
A particular area which was handled crudely below was the “relevance probability”, meaning: given that an event in the reference class occurred (e.g. given that nuclear weapons were used) what’s the probability that this event actually had an impact on the thing being modelled (e.g. what’s the probability of it interrupting the theory of change for deworming).
Table 7:
| Risk | Contribution to discount rate | Comments/sources |
| Nuclear weapons usage | 0.5% | 1.1% chance of a nuclear exchange per year (source: Rethink Priorities) ✖ relevance probability (i.e. probability that such an exchange would affect low income countries materially enough to undo the intervention being modelled). A relevance probability of 50% was chosen here, somewhat arbitrarily. |
| Other conflict/ severe geopolitical instability | 0.5% | This was estimated very quickly. Given that there are currently more than 35 armed conflicts in Africa, it seems reasonable to assume the per annum risk of conflict or other severe geopolitical event is higher than 0.5%; this was then somewhat arbitrarily adjusted down to ensure that it focuses on events severe enough to interrupt/undo the intervention being modelled. |
| Pandemics | 1% | This study from Duke University found the probability of a pandemic with similar impact to COVID-19 is about 2% in any year. I did not have time to review this study carefully. The relevance probability was set as 50%, fairly arbitrarily. |
| AI risk | 0.1% | This was derived by taking the estimate of 10% chance of an AI-related extinction event in the century (source: The Precipice, Toby Ord) and converting it to an annual figure.
|
| Other | 0.2% | Figure chosen slightly arbitrarily |
| Total | 2.3% | Because of time constraints, several components of this estimate were done quickly, and would likely be improved if assessed with more time available. |
Appendix 3: Elasticity of marginal utility of consumption
This review of values of the elasticity of marginal utility of consumption suggests that GiveWell’s choice of 1.59 is not unreasonable.

Source: p7 of Theory and Practice in the Choice of Social Discount Rate for Cost-Benefit Analysis: A survey
Here are some further references for anyone interested in researching this area further:
Appendix 4: The “temporal uncertainty” component of the discount rate is arguably being applied inconsistently between interventions
We believe this could be modelled by
We have not had the capacity to create a version of one of these models, but we would not be surprised if it were not material.
Thanks for your entry!