GiveWell’s Charity Recommendations Require Taking a Controversial Stance on Population Ethics

Cross-posted to my blog.

GiveWell claims that the Against Malaria Foun­da­tion (AMF) is about 10 times as cost-effec­tive as GiveDirectly. This en­tails un­usual claims about pop­u­la­tion ethics that I be­lieve many peo­ple would re­ject, and ac­cord­ing to other plau­si­ble views of pop­u­la­tion ethics, AMF looks less cost-effec­tive than the other GiveWell top char­i­ties.

GiveWell’s Im­plicit Assumptions

A GiveWell-com­mis­sioned re­port sug­gests that pop­u­la­tion will hardly change as a re­sult of AMF sav­ing lives. GiveWell’s cost-effec­tive­ness model for AMF as­sumes that sav­ing one life cre­ates about 35 qual­ity-ad­justed life years (QALYs), and uses this to as­sign a quan­ti­ta­tive value to the benefits of sav­ing a life. But if AMF causes pop­u­la­tions to de­cline, that means it’s ac­tu­ally re­mov­ing (hu­man) QALYs from the world; so you can’t jus­tify AMF’s pur­ported cost-effec­tive­ness by say­ing it cre­ates more happy hu­man life, be­cause it doesn’t.

You could in­stead jus­tify AMF’s life-sav­ing effects by say­ing it’s in­her­ently good to save a life, in which case GiveWell’s cost-effec­tive­ness model shouldn’t in­ter­pret the value of lives saved in terms of QALYs cre­ated/​de­stroyed, and should in­clude a term for the in­her­ent value of sav­ing a life.

GiveWell claims that AMF is about 10 times more cost-effec­tive than GiveDirectly, and GiveWell ranks AMF as its top char­ity par­tially on this ba­sis (see “Sum­mary of key con­sid­er­a­tions for top char­i­ties” in the linked ar­ti­cle). This claim de­pends on the as­sump­tion that sav­ing a life cre­ates 35 QALYs.

To jus­tify GiveWell’s cost-effec­tive­ness anal­y­sis, you could say that it is good to cause ex­ist­ing peo­ple to live longer, but it is not bad to pre­vent peo­ple from ex­ist­ing. (Sean Con­ley of GiveWell says he and many other GiveWell staffers be­lieve this.)

In par­tic­u­lar, you’d have to as­sume that:

  1. It’s valuable to cause a cur­rently-ex­ist­ing per­son to live longer, but not harm­ful to pre­vent peo­ple from ex­ist­ing.

  2. The value of caus­ing a cur­rently-ex­ist­ing per­son to live longer de­rives en­tirely from the ad­di­tional life-years lived.

(Note that (2) im­plies that sav­ing a life does not have value in it­self, but that the value de­rives from caus­ing a per­son to get to live ad­di­tional life-years.)

I will re­fer to this view as the GiveWell view, al­though I do not be­lieve all GiveWell em­ploy­ees would nec­es­sar­ily en­dorse these two as­sump­tions. Rather, I mean that this view im­plic­itly fol­lows from the way GiveWell calcu­lates AMF’s cost-effec­tive­ness. GiveWell’s recom­men­da­tions in­clude many in­puts other than ex­plicit cost-effec­tive­ness calcu­la­tions, but many donors nonethe­less heav­ily rely on the claim that AMF is about 10 times more cost-effec­tive than GiveDirectly.

GiveWell’s cost-effec­tive­ness calcu­la­tions not only claim that pre­vent­ing deaths from malaria is good, but that its good­ness comes in pro­por­tion to the num­ber of ad­di­tional life-years lived by peo­ple who oth­er­wise would have died from malaria.

A Paradox

This po­si­tion leads to counter-in­tu­itive and pos­si­bly con­tra­dic­tory con­clu­sions. Let’s look at a thought ex­per­i­ment to see why.

You can choose be­tween three pos­si­ble acts. When you perform act1 A, a child (let’s call her Afiya) is born, gets malaria, and dies. Un­der act B, you cause Afiya not to be born. Ac­cord­ing to as­sump­tion (1), this act is not worse than A. Stan­dard per­son-af­fect­ing view says that it is not wrong to cause some­one to ex­ist whose life is net pos­i­tive, so A is not worse than B. Un­der act C, you cause Afiya to be born and pre­vent her from get­ting malaria. This beats act A ac­cord­ing to (2), and is not bet­ter than act B ac­cord­ing to (1). Thus, A = B, B ≥ C, and C > A. But this cre­ates a con­tra­dic­tion: B > A and B = A.

This same con­cern ap­plies to any per­son-af­fect­ing view. I be­lieve the most plau­si­ble re­s­olu­tion of this para­dox comes from Roberts (2003)2: It is not wrong to fail to bring some­one into ex­is­tence, but it is wrong to bring some­one into ex­is­tence who will be less happy than they could be. So you ought not al­low Afiya to be born and con­tract malaria when you could have pre­vented her from ex­ist­ing or caused her not to con­tract malaria.In other words, B > A, C > A, and B ≥ C.

For a stan­dard per­son-af­fect­ing view, this suc­cess­fully avoids cre­at­ing a con­tra­dic­tion (al­though I and many oth­ers find it un­satis­fy­ing). How­ever, the GiveWell view has ad­di­tional prob­lems that a stan­dard per­son-af­fect­ing view does not, be­cause the GiveWell view in­cludes the ad­di­tional as­sump­tion that the value of caus­ing a cur­rently-ex­ist­ing per­son to live longer de­rives en­tirely from the ad­di­tional life-years lived.

By as­sump­tion (2), since Afiya gains an ad­di­tional 35 QALYs when she doesn’t get malaria, we can say that C = A + 35 (that is, act C is “35 QALYs” bet­ter than act A). Since B ≥ C, that means B ≥ A + 35. In other words, if we pre­vent Afiya from ex­ist­ing, that’s at least 35 QALYs bet­ter caus­ing Afiya to be born and then let­ting her die of malaria.

This might not ap­pear ex­plic­itly con­tra­dic­tory, it nonethe­less comes across as very weird. How can act B be worth 35 life-years “bet­ter” than act A if it in fact con­tains fewer happy life-years?

In short, to avoid in­ter­nal con­tra­dic­tion, the GiveWell view must claim caus­ing Afiya not to be born is worth at least 35 QALYs “more” than caus­ing her to be born and then die of malaria. (There ex­ist a few other pos­si­bil­ities, such as that tran­si­tivity does not hold for or­der­ings over acts, but I find these suffi­ciently im­plau­si­ble that I will not ad­dress them here.)

A cou­ple of com­menters have pro­posed that a child’s life may be net nega­tive on bal­ance when she dies of malaria at a young age. This still does not avoid the prob­lem with the GiveWell view where the differ­ence be­tween the child’s ex­is­tence and nonex­is­tence ex­actly equals the value of the ad­di­tional life-years she would have if she didn’t get malaria. And, more sig­nifi­cantly, this does not re­solve the ad­di­tional prob­lems with the GiveWell view de­scribed in the next sec­tion.

Big­ger Problems

We can ac­tu­ally make a stronger claim than this: the GiveWell view ap­pears to ex­plic­itly con­tra­dict Roberts’s re­s­olu­tion to the per­son-af­fect­ing view. Show­ing this gets a bit com­pli­cated, so bear with me.

In this sce­nario, there are two chil­dren, Afiya and Brian, who might po­ten­tially get malaria. You may take one of four pos­si­ble acts.

  • Act A: Afiya gets malaria and dies. Brian is then born, gets malaria, and dies.

  • Act B: You pre­vent Afiya from get­ting malaria. Brian is still born and dies of malaria.

  • Act C: You pre­vent Afiya from get­ting malaria and Brian is never born.

  • Act D: Both Afiya and Brian are born, and you pre­vent them both from get­ting malaria.

3

Ac­cord­ing to the GiveWell view, A < B and B < D.

Un­der the stan­dard per­son-af­fect­ing view, this cre­ates the same para­dox as our pre­vi­ous thought ex­per­i­ment. But even Roberts’s re­s­olu­tion does not work here for the GiveWell view. Roberts’s re­s­olu­tion en­tails that C = D and there­fore B < C. But GiveWell’s cost-effec­tive­ness calcu­la­tions for AMF im­plic­itly en­dorse B = C. When we donate to AMF and save an Afiya, stan­dard cost-effec­tive­ness es­ti­mates for AMF don’t give any re­gard to whether a Brian ex­ists, even though un­der Roberts’s view we should pre­fer Brian not to ex­ist. If donat­ing to AMF looks like act B, then sav­ing a life is worth 35 QALYs; but if it looks like act C, then sav­ing a life is worth 70 QALYs. Similarly, acts C and D count for the same un­der Roberts’s re­s­olu­tion, even though GiveWell would count act D as twice as good as act C.

The GiveWell view as­sumes that the benefits of pre­vent­ing Afiya from get­ting malaria come en­tirely from caus­ing her to live longer. The GiveWell view is in­differ­ent about whether sav­ing Afiya counts as act B, C, or D. In other words, B = C = D. But at the same time, D > B be­cause the GiveWell view wants us to save Brian. This is not just con­tro­ver­sial—it’s in­ter­nally in­con­sis­tent.

Some philoso­phers still ac­cept some­thing like the GiveWell view in spite of its prob­lems4. But even if we ad­mit the plau­si­bil­ity of the GiveWell view, I do not be­lieve GiveWell should claim that AMF is 10 times more cost-effec­tive than GiveDirectly. This claim re­lies on fairly spe­cific as­sump­tions about pop­u­la­tion ethics that many po­ten­tial donors will re­ject. If donors un­der­stood these as­sump­tions, I ex­pect that many of them would pri­ori­tize their dona­tions differ­ently. For ex­am­ple, last year when Stan­ford Effec­tive Altru­ism was con­sid­er­ing mak­ing dona­tions to char­ity, we preferred the Schis­to­so­mi­a­sis Con­trol Ini­ti­a­tive over AMF be­cause we be­lieved that GiveWell gave too much sig­nifi­cance to the “GiveWell view” of pop­u­la­tion ethics and not enough to the to­tal view. Had we not un­der­stood the sig­nifi­cance of pop­u­la­tion ethics in GiveWell’s recom­men­da­tion, we might have pri­ori­tized AMF even though our ac­tual be­liefs about pop­u­la­tion ethics would im­ply that the other top char­i­ties look bet­ter.

At­tempted Resolutions

The GiveWell view has un­de­sir­able im­pli­ca­tions for pop­u­la­tion ethics. This gives rise to two con­cerns: (i) donors may end up fol­low­ing recom­men­da­tions that con­tra­dict their views about pop­u­la­tion ethics; and (ii) GiveWell is prob­a­bly in fact wrong about its cost-effec­tive­ness calcu­la­tions, which count for a large part of its fi­nal char­ity recom­men­da­tions. Con­cern (i) sug­gests that GiveWell should make its as­sump­tions about pop­u­la­tion ethics more ex­plicit, and con­cern (ii) sug­gests that GiveWell should change its cost-effec­tive­ness es­ti­mates and pos­si­bly its top char­ity recom­men­da­tions.

I be­lieve that even peo­ple who en­dorse the GiveWell view should agree with con­cern (i). On con­cern (ii), I see two strong can­di­dates for how to re­solve the prob­lems with the GiveWell view:

  1. Re­ject as­sump­tion (1)–prob­a­bly by adopt­ing a to­tal view, which claims that it is good rather than neu­tral to cre­ate new happy peo­ple.

  2. Re­ject as­sump­tion (2), per­haps by claiming that pre­vent­ing deaths has value be­yond just the ad­di­tional life years cre­ated. This would bring us closer to some­thing like prefer­ence util­i­tar­i­anism with a per­son-af­fect­ing view of pop­u­la­tion ethics.

Let’s see if these pro­duce similar coun­ter­in­tu­itive con­clu­sions to the GiveWell view, and how we would pri­ori­tize char­i­ties differ­ently if we adopted them.

First Resolution

If we fol­low the first re­s­olu­tion, the value of AMF de­rives from its global effect on in­creas­ing hap­piness and re­duc­ing suffer­ing. Be­cause AMF hardly changes hu­mans’ lifes­pans, it does not have a clear benefi­cial effect for hu­mans (and it’s un­clear whether cre­at­ing more happy hu­man life benefits or harms sen­tient life in gen­eral). So un­der this re­s­olu­tion, AMF looks much less effec­tive. I per­son­ally pre­fer this view, as do many peo­ple at Stan­ford Effec­tive Altru­ism (as men­tioned above); and as do many philoso­phers for that mat­ter5. Thus, I sub­stan­tially dis­count the life-sav­ing benefits of AMF when pri­ori­tiz­ing GiveWell top char­i­ties, and I be­lieve GiveWell ought to as well.

Se­cond Resolution

The sec­ond re­s­olu­tion may have even worse im­pli­ca­tions than the GiveWell view. Sup­pose we re­ject as­sump­tion (2) by claiming that pre­vent­ing death has ex­tra value be­yond its effect on some­one’s lifes­pan. (Let’s call this the up­dated per­son-af­fect­ing view.) Re­turn­ing to our ini­tial thought ex­per­i­ment, the differ­en­tial be­tween act A (in which a child is born and dies of malaria) and act C (in which a child is born and pre­vented from get­ting malaria) ap­pears big­ger un­der the up­dated per­son-af­fect­ing view than un­der the GiveWell view. But we still can­not say that we pre­fer act C to act B, which means now in­stead of hav­ing B ≥ A + 35, we have B > A + 35.

We can avoid this prob­lem by claiming that the life-years added by act C have sub-lin­ear value with the length of life. This view ap­pears more plau­si­ble, but it means we have no easy way to di­rectly com­pare AMF against GiveDirectly (or against other char­i­ties that serve to im­prove peo­ple’s lives rather than pre­vent them from dy­ing). We would need to cre­ate some new way to trade off pre­vent­ing deaths against im­prov­ing lives, so our cost-effec­tive­ness calcu­la­tions for AMF would look sub­stan­tially differ­ent.

Conclusion

Cost-effec­tive­ness es­ti­mates for GiveWell top char­i­ties, and there­fore GiveWell’s recom­men­da­tions6, de­pend heav­ily on con­tro­ver­sial ques­tions about pop­u­la­tion ethics. I be­lieve GiveWell takes an in­cor­rect stance here, and it ought to give more weight to the to­tal view. But more im­por­tantly, donors should be aware of how ques­tions of pop­u­la­tion ethics af­fect the ex­pected value of differ­ent in­ter­ven­tions.

To GiveWell’s credit, it has writ­ten about how its cost-effec­tive­ness es­ti­mates re­quire mak­ing judg­ment calls. How­ever, if I am read­ing them cor­rectly, the cost-effec­tive­ness analy­ses done by differ­ent GiveWell em­ploy­ees all uniformly as­sume that sav­ing one life with malaria nets has the same value as adding ~35 QALYs, which means GiveWell’s cost-effec­tive­ness spread­sheet en­forces a sin­gle judg­ment call on pop­u­la­tion ethics (i.e., what I have called the GiveWell view). GiveWell’s list of key as­sump­tions on AMF does not in­clude any­thing about pop­u­la­tion ethics.

A large part of the case for AMF rests on the fact that it ap­pears highly cost-effec­tive when you make cer­tain as­sump­tions about pop­u­la­tion ethics. Donors who do not make these as­sump­tions, or who have sub­stan­tial un­cer­tainty about them, should con­sider which char­i­ties look best on their own view of pop­u­la­tion ethics.

Er­rata: I origi­nally used DALYs in­stead of QALYs be­cause GiveWell’s cost-effec­tive­ness es­ti­mates use DALYs. I was un­aware that DALYs have the op­po­site sign of QALYs. I changed this post to use QALYs in­stead of DALYs.

Notes

  1. I origi­nally had this eval­u­at­ing world states rather than acts, but Ei­tan Fischer pointed out that the world-ori­en­ta­tion cre­ates more sub­stan­tial prob­lems than the act-ori­en­ta­tion.

  2. Roberts, M. A. (2003). Is the per­son-af­fect­ing in­tu­ition para­dox­i­cal?. The­ory and De­ci­sion, 55(1), 1-44. http://​​link.springer.com/​​ar­ti­cle/​​10.1023%2FB%3ATHEO.0000019052.80871.b3

  3. As you can clearly see, I am not a graphic de­signer.

  4. All eth­i­cal views have coun­ter­in­tu­itive parts; it’s just a mat­ter of choos­ing which of your in­tu­itions you want to vi­o­late.

  5. The most com­mon ob­jec­tion to the to­tal view is the re­pug­nant con­clu­sion, al­though I do not be­lieve that the re­pug­nant con­clu­sion is in fact re­pug­nant.

  6. This is not to say that recom­men­da­tions are en­tirely de­ter­mined by cost-effec­tive­ness es­ti­mates, but that such es­ti­mates are a ma­jor in­put.