This was a shallow review conducted by SoGive aiming to quantitatively explore the extent to which tackling poverty reduces the incidence of conflict/war. As this was a shallow review its conclusions are tentative.
A model looking at the link between poverty and conflict risk found the following:
- If you are donating to tackle poverty, how material is the flowthrough effect of reducing the risk of conflict? Ignoring tail risk, the materiality was small; i.e. the cost per conflict-related death averted was expensive (in the tens of millions of dollars).
- However if the model linking poverty and conflict can equally be applied in a nuclear context (and this is a big if) then the flowthrough effects alone may outweigh the direct effects; i.e. the cost per conflict-related death averted was competitive with GiveWell-recommended interventions (c$10,000). So tackling poverty in a potentially nuclear context might do more good from avoiding nuclear conflict than from directly helping people.
Does this mean that more effort should be put into tackling poverty in nuclear contexts (e.g. India/Pakistan)? Maybe. This review was too shallow to form confident conclusions on this, but sets out further research to perform on this topic (including the risks of moral hazard).
SoGive is an organisation conducting analysis to help donors make better donation decisions.
This was a shallow review, meaning that it was time-limited. So several elements of the analysis which would have been considered more carefully in a more rigorous review have been glossed over here. This is partly to manage our time, and partly because of a “lean analysis” approach: publishing sooner helps us to gather feedback sooner. Compared to other shallow reviews we have conducted, this was particularly shallow relative to the complexity of the topic. So it’s intended more as a basis for further discussion and improvement, and not intended as the final word on the topic.
Section 1 sets out the high-level structure of the model
Section 2 includes some discussion of the findings.
Section 3 sets out the assumptions which we think require further analysis.
1. High level structure of the model
The model was based on a putative intervention to eliminate poverty. The intervention was GiveDirectly (giving cash directly to the poor) to all those below the poverty line in Kenya for a non-nuclear context, and similarly again focusing on India/Pakistan for the nuclear context.
· GiveDirectly was chosen because of a moderately high level of confidence that the intervention will be effective at lifting people out of poverty.
· Kenya was chosen because it is a roughly average sized developing world country with relatively good levels of data availability.
· India/Pakistan were chosen as an example of nuclear states with tensions and risks of conflict and where there is poverty.
The assumption about the extent to which the probability of conflict varies as wealth/GDP improve comes from Paul Collier’s work. In his book The Bottom Billion; he writes “Each percentage point added to the growth rate knocks off a percentage point from this [war] risk”.
Collier’s model is not the only one which provides a quantitative link between poverty and risk of conflict. Another model by Humphries considers the probability of conflict to be inversely proportional to GDP. Under this model, the impact of tackling poverty appears to be around one order of magnitude (i.e. roughly a factor of ten) less impactful. This does not change the two main conclusions (namely that excluding tail risk, the flowthrough impact of tackling poverty is small, but that where the conflict could be nuclear, the flowthrough benefits of avoiding a nuclear war could be competitive (or at least almost competitive) with the more direct/immediate benefits). This still relies on the assumption set out in section 3 that the models applicable for “normal” war also apply to nuclear war.
Then the number of deaths per war is modelled on a basis ignoring tail risk (i.e. based on the average number of deaths since 1989). It’s then remodelled taking into account a small probability of an extreme scale of war (i.e. including tail risk), based on pre-existing estimates of the scale of and probability of nuclear wars. As discussed in section 3, an important assumption is that the same model can be used in this nuclear context.
The expected number of conflict-related deaths averted is compared with the cost of achieving the intervention.
This is then based on our usual thresholds, based on GiveWell-recommended charities.
First the threshold (for comparison purposes):
· c $2,000 per life saved would count as a “good” cost per life saved equivalent (as described on the SoGive cost-effective benchmarks page, which is based on GiveWell’s models)
On a basis excluding tail risk:
· Cost per conflict-related death averted as a result of reducing poverty > $10 million
On a basis including tail risk:
· Cost per conflict-related death averted as a result of reducing poverty = c$10,000.
· For comparison, the SoGive cost-effective benchmarks page shows that, according to the November 2019 GiveWell cost-effectiveness analysis model, GiveDirectly’s cost per life-saved equivalent was around $29,000.
This tentatively suggests that in a geography where nuclear war is a realistic possibility, the flowthrough effects of tackling poverty may outperform the immediate effects.
As this was a quick, shallow review, the considerations set out in the following section should be given some more thought before these results are treated as final.
For the modelling of reducing poverty, the model assumes that each poor person receives a cash injection of $1,000, which earns an investment return of 10% (which is a somewhat uncertain assumption, taken from GiveWell). It also assumes that the $1,000 is spent down over 10 years. This results in an improvement in GDP of $200 per capita, on top of the current GDP per capita of just over $1,700 (source: World Bank) leading to an improvement of growth of just over 4%.
The model didn’t consider the costs involved in implementing the GiveDirectly intervention. The proportion of each donation going through to the beneficiary is fairly high (>80%: source: GiveWell cost-effectiveness analysis). Even if implementing this intervention at scale required more overheads (i.e. economies of scale did not, for some reason, apply) the model being applied here is sufficiently crude that modelling this level of detail would constitute spurious accuracy.
Note that the GiveDirectly intervention is assumed to be effective based on assessments performed on small-scale trials. If the intervention were performed at the scale assumed in this model, further considerations should be taken into account, such as the possibility of Dutch disease. (see, e.g. https://www.tandfonline.com/doi/abs/10.1080/10168738800000020, which examines this in the context of foreign borrowing). Dutch disease was not modelled here. Note that the putative implementation of GiveDirectly’s intervention at the level of the whole of Kenya is a thought experiment to help us to model effects on a consistent basis. In reality, we expect this thinking to inform donations happening at a smaller scale.
Conversely, it seems plausible to imagine that another intervention may be effective at lifting people out of poverty more cost effectively (see e.g. https://forum.effectivealtruism.org/posts/bsE5t6qhGC65fEpzN/growth-and-the-case-against-randomista-development).
The improvement in growth of just over 4% meant that the probability of war was also reduced by just over 4% over the coming 5 years. The present value of this was modelled as 0.16 wars averted, in expectation.
Intuitively, only 0.16 wars averted for a wide-scale, expensive programme tackling poverty at scale might sound like a small impact. This intuition is formalised by converting this into a cost per life saved (or rather cost per battle-death averted), which will be described shortly. (Note we understand the battle-related death data used to already include civilian casualties.)
Note that there are other implications of war apart from lives lost – e.g. war may bring about non-fatal casualties or damage to infrastructure, which may impact on economic productivity. Our shallow review did not find good global data on non-fatal casualties, however a quick glance at the Wikipedia page on US military casualties of war (https://en.wikipedia.org/wiki/United_States_military_casualties_of_war) found that the number of non-fatal casualties appears to be not substantially more than the number fatalities. (Ideally we would use data comparing casualties to deaths in a not-specific-to-the-military context, and a not-specific-to-the-US context, however we did not find this in our time-constrained review) A fuller model would estimate the number of non-fatal casualties, estimate a suitable average DALY (disability-adjusted life year) weighting for a war casualty, and model this as well. As we are looking at this in terms of orders of magnitude, these issues seem unlikely to significantly move the dial, or materially change the conclusions.
This was then converted into a cost per life saved by calculating the average number of lives lost in a conflict based on data from the UCDP or Uppsala Conflict Data Programme (Pettersson, Therese; Stina Högbladh & Magnus Öberg, 2019. Organized violence, 1989-2018 and peace agreements, Journal of Peace Research 56(4)) which found that the average number of battle deaths per conflict over the period 1989-2018 was just under 8,000. The 8,000 death severity was taken as a measure of the badness of a typical conflict ignoring tail risk.
For the including-tail-risk figure, the assumptions around the scale of and probability of nuclear war are taken from the Global Catastrophic Risks Survey 2008 (https://www.fhi.ox.ac.uk/reports/2008-1.pdf). This choice is based on the review conducted by Luisa Rodriguez which quoted this survey. It has been adjusted down by a factor of ten, which broadly captures the idea that the intervention covers one specific area (e.g. India/Pakistan)
Importantly, the model gave no credit to people not born yet. If we thought that the extinction scenario should also factor those people in, the model would give substantially more weight to the tail risk.
The model includes important assumptions about the modelling of the link between poverty and conflict, which is the heart of this exercise. In particular, the first variant is based on the Collier model. This appears to be more geared towards civil war, and is based on modelling linked to the rebel greed hypothesis. Our review has been too shallow to consider the basis for the Collier model in any depth, however it appears at first glance that linking this model to “normal”-sized wars (e.g. those with a death count of c 8000 lives) seems reasonable, however our shallow review of the Collier model has not been rigorous enough to determine whether this is appropriate for a nuclear war as well.
It seems reasonable to imagine that Collier may not have been thinking of this type of conflict, but that a link between poverty and a propensity to behave in a bellicose fashion likely still exists, even when it comes to international conflicts. And it may be that the models which apply to “normal” conflict are also (roughly) transferable to nuclear conflict – the key assumption being made here is that the quantum of that link is (in very very rough, ballpark terms) similar to that used in the Collier model. (Or similar to the Humphreys model in variant two). Mechanisms by which poverty could lead to nuclear conflict include poverty being linked to weak governance at state level, which may make a state more bellicose, or a poverty-stricken country being more prone to the “rebel greed” hypothesis that Collier sets out. These thoughts would need further elaboration in a less shallow review.
One more area which would need to be considered more carefully is moral hazard. Does providing support to geographies that have nuclear capability provide perverse incentives? This alone may be sufficient to quash this line of thought. A fuller consideration of this would need to consider how realistic the moral hazard is (i.e. how realistic is it for a state considering developing nuclear capability to be influenced by the thought of receiving more aid). It would also need to consider which donors are being influenced by this thinking (e.g. if DfID or USAID had an explicit policy of providing more aid to nuclear states, this might be different from a single philanthropist carefully considering which areas of poverty are likely to lead to nuclear conflict, and not stating this in their rationale for their grant making).
If this shallow review were to be expanded upon and tackled more carefully, the elements needing more attention are all the assumptions listed above, with a particular focus on a deeper understanding of the Collier and Humphreys models, as well as any other models. This includes not only critiquing the models to have greater confidence in their robustness, but also understanding whether the dynamics behind nuclear war have enough similarities or areas of crossover to warrant these models being leveraged in the nuclear context, or, alternatively, replacing them with models which are suitable for that context.
The model itself can be found here: https://docs.google.com/spreadsheets/d/1FVMibJXxM72QOQBelR0goTUTPWSMOJ2pCoSt4sOIofs/edit#gid=0