Response to a Dylan Matthews article on Vox about bipartisanship

This is a dis­pro­por­tionately long take­down of an op-ed. It is rele­vant for judg­ing Vox’s EA jour­nal­ism and figur­ing out EA norms on poli­tics and jour­nal­ism. It also has a bit of in­for­ma­tion on Amer­i­can poli­tics and cul­ture if you are in­ter­ested in that.

This ar­ti­cle by Dy­lan Matthews (“The ‘why can’t we all just get along’ the­ory of poli­tics”) was pub­lished two weeks ago un­der the “Fu­ture Perfect” column of EA-as­so­ci­ated jour­nal­ism, funded by the Rock­efel­ler Foun­da­tion. Most of these ar­ti­cles have been good but this one is cer­tainly the worst out of all that I have seen (n=25 or so, from mul­ti­ple writ­ers) and I be­lieve it has nega­tive ex­pected value.

The ar­ti­cle gives a ba­sic line of ar­gu­ment that has been rel­a­tively com­mon in Western poli­ti­cal cul­ture since 2016 or so—that we don’t need to try to get along with our poli­ti­cal op­po­nents. He nom­i­nally ar­gues for this in cost-benefit terms. In­deed, the prac­ti­cal cost of mak­ing friends with ex­tra peo­ple is perfectly real and rele­vant:

For av­er­age cit­i­zens and vot­ers, [go­ing out of your way to make friends with poli­ti­cal op­po­nents] is an­other bur­den to add to the list af­ter work, schlep­ping the kids to and from school, tak­ing care of el­derly fam­ily mem­bers, and at­tend­ing PTA and church/​syn­a­gogue/​mosque meet­ings, etc.
What the call for cross-par­ti­san friend­ships asks peo­ple to do, es­sen­tially, is to make an al­tru­is­tic sac­ri­fice of time, per­haps money, and definitely emo­tional en­ergy, in an at­tempt to heal our poli­tics.

And that effort could be spent more pro­duc­tively. This makes perfect sense. Yet while this claim is defen­si­ble and rea­son­able, the ar­ti­cle heav­ily pushes a much more poorly sup­ported im­plicit point of view that Matthews’ side of poli­tics is un­con­tro­ver­sially the cor­rect side, that there is lit­tle or noth­ing to learn from oth­ers, and that the oth­ers are morally re­pug­nant. In do­ing so, it goes well be­yond the nom­i­nal ar­gu­ment for shrewd­ness and dis­plays ac­tive an­tag­o­niza­tion in­stead.

First, Matthews begs the ques­tion by fram­ing the main goal of bi­par­ti­san friend­ship as “at­tempt[ing] to per­suade par­ti­ci­pants of cer­tain spe­cific, so­cially benefi­cial be­liefs” and pre­sent­ing his own views as the de­fault good ones. This at­ti­tude is clearly an ob­sta­cle to hav­ing a mu­tu­ally re­spect­ful friend­ship. Now we can cer­tainly rank be­liefs in terms of their ex­pected so­cial benefit. But to ex­plic­itly push this at­ti­tude and our own cre­dences into our friend­ships with poli­ti­cal op­po­nents in­stantly de­stroys the nec­es­sary premise for any such friend­ship to achieve its pri­mary teloses—mak­ing the poli­ti­cal sys­tem work bet­ter for mul­ti­ple points of view, or learn­ing more from your op­po­nent.

Matthews also com­mits a ba­sic de­ci­sion-the­o­retic er­ror here in ig­nor­ing the ro­bust­ness of our be­liefs and the value of in­for­ma­tion. If our own poli­ti­cal be­liefs on any im­por­tant is­sue are not very ro­bust, then high value can be had by in­ves­ti­gat­ing un­known ev­i­dence to re­fine our ex­pected value es­ti­mates. Matthews is not only as­sum­ing that his be­liefs are benefi­cial, he is as­sum­ing that other peo­ple’s ar­gu­ments lack enough po­ten­tial merit to jus­tify listen­ing to them.

As for the be­liefs in ques­tion, Matthews says the fol­low­ing:

In­so­far as in­di­vi­d­ual be­liefs are de­form­ing our poli­tics, the be­liefs that do so the worst in­volve bi­gotry — es­pe­cially, in the Amer­i­can case, racist sen­ti­ment.

As we shall see, Matthews’ jus­tifi­ca­tion for this as­sump­tion is neg­ligible. Of course it is by no means an un­com­mon or un­sup­port­able as­sump­tion, but for Matthews to ca­su­ally as­sert it with such high con­fi­dence and such poor jus­tifi­ca­tion reaches a crim­i­nally bad level of epistemic hy­giene.

Matthews sup­ports it only with a hy­per­link to a story by an­other Vox jour­nal­ist, Ger­man Lopez, who ar­gues that “Trump won be­cause of racial re­sent­ment.” Yet that ar­ti­cle has ma­jor prob­lems, both in its own right and in terms of the val­idity of its ev­i­dence for the pur­pose of Matthews’ cita­tion.

Its first source, a Wash­ing­ton Post ar­ti­cle by Fowler et al has data show­ing that Trump vot­ing was pre­dicted by white vuln­er­a­bil­ity, and white vuln­er­a­bil­ity is in turn pre­dicted by “racial re­sent­ment”.

I do not see a pa­per show­ing Fowler et al’s ac­tual anal­y­sis (!), but the sur­vey data they used sug­gests a pos­si­ble prob­lem of sound­ness, albeit a rel­a­tively minor one that we may for­give Vox for over­look­ing. Their race at­ti­tude ques­tions seem to be from Q20 to Q51, the ma­jor­ity of the sur­vey. Their ‘white vuln­er­a­bil­ity’ ques­tions are, ac­cord­ing to the WaPo piece, Q29-Q31. But Q28A, Q28B, and Q32-Q35 can just as plau­si­bly be con­strued as ques­tions about white vuln­er­a­bil­ity, and noth­ing in the sur­vey sug­gests pre­s­e­lec­tion of Q29-Q31. It’s pos­si­ble that the au­thors cher­ryp­icked here, so that the idea of white vuln­er­a­bil­ity and the ques­tions used to con­sti­tute it were cho­sen from other equally racial al­ter­na­tives ex post, mean­ing that the cor­re­la­tion would prob­a­bly be un­rep­re­sen­ta­tively high as well as over­es­ti­mated (analo­gous to the Op­ti­mizer’s Curse). Similarly, it’s not ap­par­ent whether the au­thors used all of Q20-Q28 + Q32-Q51 to mea­sure racial re­sent­ment or just some of them; again there is po­ten­tial here for ex post fudg­ing.

Look­ing up racial re­sent­ment scales on­line, I find this un­der­grad­u­ate’s ar­ti­cle in a non­par­ti­san poli­ti­cal jour­nal sum­ma­riz­ing and com­ment­ing on re­sults from two stud­ies which show that such scales can mis­take gen­eral con­ser­vatism for racial re­sent­ment. Many of Fowler et al’s race ques­tions seem to be rele­vantly similar, which sug­gests that they are not valid mea­sures of racial re­sent­ment. More in­ter­est­ingly, Fowler et al pre­sum­ably did not find that racial re­sent­ment pre­dicted Trump vot­ing; they merely found that it pre­dicted the white vuln­er­a­bil­ity that pre­dicted Trump vot­ing. If they had found a sig­nifi­cant di­rect con­nec­tion, surely they would have men­tioned that. The omis­sion sug­gests that the con­nec­tion is specious, e.g. re­sent­ment only pre­dicted vuln­er­a­bil­ity among groups who were vot­ing for Trump any­way or some­thing of the sort. I think any good so­cial sci­en­tist would agree that pre­dic­tion is not tran­si­tive across so­cial phe­nom­ena when you are con­trol­ling for con­found­ing vari­ables.

So what do Fowler et al re­ally tell us? That Trump vot­ing is pre­dicted by Q29-Q31, namely the be­liefs that 1) “through no fault of their own, whites are eco­nom­i­cally los­ing ground to­day com­pared to other racial and eth­nic groups,” that 2) “dis­crim­i­na­tion against whites has be­come as big a prob­lem as dis­crim­i­na­tion against blacks and other minori­ties”, and that 3) minori­ties over­tak­ing whites by 2050 would weaken the coun­try.

The sec­ond study cited by Lopez is Schaf­fner et al, which shows that Trump­ism is strongly pre­dicted by the com­bined met­ric of 4) re­jec­tion of the claim that “white peo­ple in the U.S. have cer­tain ad­van­tages be­cause of the color of their skin,” 5) en­dorse­ment of the claim that “racial prob­lems in the U.S. are rare, iso­lated situ­a­tions,” and 6) not be­ing “an­gry that racism ex­ists.” Other ev­i­dence cited by Lopez shows that Trump vot­ers were vuln­er­a­ble to sub­con­scious poli­ti­cal bias by racial trig­gers.

Still more ar­gu­ments are given by Lopez but they are go­ing to be ig­nored here be­cause even if we take them at face value they do not give clear sup­port to Matthews’ as­ser­tion that racist sen­ti­ment among vot­ers elected Trump. Now, as­sum­ing that the ba­sic re­sults of these stud­ies are cor­rect and as­sum­ing that Lopez has not cher­ryp­icked ev­i­dence to fit his point of view, let’s stop and re­view Matthews’ po­si­tion.

He has not in­cluded any ra­tio­nale for con­sid­er­ing Trump to be a par­tic­u­larly bad pres­i­dent in the first place, al­though many brief ar­gu­ments can be found scat­tered across the writ­ings of him­self and oth­ers on Vox so we can per­haps leave that aside.

He com­pletely ig­nores the harms that can be caused by all sorts of other be­liefs. Even with the as­sump­tion of Trump be­ing a par­tic­u­larly bad pres­i­dent, it is non-ob­vi­ous that the be­liefs that elected him are more harm­ful (es­pe­cially on a per-per­son ba­sis) than be­liefs which cre­ate other poli­ti­cal prob­lems. For in­stance, con­sider is­sues of farm leg­is­la­tion, obli­ga­tions to fu­ture peo­ple, for­eign aid, and sup­port for Effec­tive Altru­ism.

And to sum­ma­rize the above re­view of Lopez’s ar­ti­cle, all that we can say with rigor about the sen­ti­ment at play among Trump vot­ers is that it in­volves the six afore­men­tioned be­liefs (or maybe some sub­set) as well as im­plicit racial fears and bi­ases.

Matthews be­lieves that it is painful and costly for a minor­ity per­son to have to listen to some­one who has the above at­ti­tudes. But there are three prob­lems with this. First, it is not clear how they will make some­one un­usu­ally painful or difficult to befriend. After all, they are not suffi­cient for the or­tho­dox rigor­ous defi­ni­tion of racism, Jorge Gar­cia’s vo­li­tional ac­count, which is es­pe­cially the most rele­vant when we are talk­ing about in­ter­per­sonal re­la­tions. (Note: even if you be­lieve that the other racial ques­tions given by Fowler et al are pre­dic­tors of Trump vot­ing, most of them are similarly very weak in this re­gard, and none of them are par­tic­u­larly strong.) Go­ing by per­sonal ex­pe­rience, I am Ar­me­nian, yet if I meet a Turk­ish per­son I will not feel un­duly emo­tion­ally trou­bled merely be­cause they are not suffi­ciently an­gry about the Ar­me­nian Geno­cide, and so on for similar analo­gies.

Se­cond, the cor­re­la­tions still do not show that Trump vot­ers have a suffi­ciently se­vere case of these traits so as to make re­la­tion­ships suffi­ciently costly. What is the thresh­old?

Third, Matthews uses the ex­am­ple of a Mus­lim woman hear­ing about a travel ban as an in­tu­ition pump to show how costly it is to befriend a con­ser­va­tive. Yet the con­ser­va­tives may only have racial sen­ti­ments with re­spect to some minor­ity groups and not oth­ers, and those of us who are mak­ing friends with Trump vot­ers may not be minori­ties at all. In that case, the force of Matthews’ ar­gu­ment dis­si­pates fur­ther. In line with his gen­er­ous in­ter­pre­ta­tion of the phrase “racist sen­ti­ment”, Matthews char­ac­ter­izes sup­port for the travel ban as “hate” for a Mus­lim woman and her fam­ily.

Matthews then ar­gues that con­flict, such as that on Twit­ter, can be benefi­cial. Matthews looks to Bruneau and Saxe as sup­pos­edly show­ing that con­flict in the form of “airing grievances” works in mak­ing peo­ple more mod­er­ate, but read­ing their study re­veals that, far from be­ing rem­i­nis­cent of Twit­ter mud­sling­ing, it ac­tu­ally ap­pears to be con­sti­tuted by the shar­ing of care­fully-writ­ten poli­ti­cal views fol­lowed by a rudi­men­tary ver­sion of the Ide­olog­i­cal Tur­ing Test (!). That sounds ex­actly like a bi­par­ti­san friend­ship to me. His sec­ond point is his per­sonal rec­ol­lec­tion of how white jour­nal­ists like him­self learned from blacks on Twit­ter. But it should be ap­par­ent by now that Dy­lan Matthews is prob­a­bly not some­one who has racist sen­ti­ments against blacks, mak­ing the anec­dote in­valid.

Matthews sug­gests that the value of con­flict pro­vides ev­i­dence against the value of bi­par­ti­san friend­ships, yet it is a weak in­fer­ence. It could plau­si­bly be the case that both friend­ships and con­flicts are use­ful, maybe in differ­ent con­texts. In­deed, the other study that he cites here clearly ac­knowl­edges that “co­op­er­a­tive be­hav­iors are so­cially and eco­nom­i­cally benefi­cial across a large va­ri­ety of con­texts.” Matthews also gives a met­anal­y­sis which says that friendly con­tact has a real but weak effect in re­duc­ing racial prej­u­dice (the lat­ter be­ing mea­sured in many ways across a va­ri­ety of stud­ies); he leaves out the more op­ti­mistic met­anal­y­sis by Pet­ti­grew and Tropp but that one is older so we can let it pass.

So, for the ques­tion of the value of bi­par­ti­san di­alogue, Matthews:

  • Ig­nored one of the main goals of bi­par­ti­san di­alogue, sys­temic changes to poli­ti­cal culture

  • Ig­nored the other main goal, the pos­si­bil­ity of learn­ing from one’s poli­ti­cal opponents

  • Did not in­clude jus­tifi­ca­tions for his view that Trump’s elec­tion had par­tic­u­larly bad consequences

  • Pro­vided no com­par­a­tive ar­gu­ment to show that Trump’s elec­tion and the rele­vant be­liefs are the most harm­ful is­sues to consider

  • Pro­vided very weak sup­port for his claim that Trump vot­ers tend to hold sen­ti­ments that will of­ten offend well-mean­ing minor­ity friends, with the rele­vant stud­ies turn­ing out to be in­ves­ti­ga­tions of at­ti­tudes which do not seem likely to cause this effect

  • Ig­nored non-minor­ity read­ers and read­ers who are of minori­ties that are not tar­geted by Trump vot­ers’ attitudes

  • Pro­vides mediocre sup­port for his claim that co­op­er­a­tion and con­tact lack suffi­cient effi­cacy to pre­fer them over conflict

Then, for the EA com­mu­nity it­self, this ar­ti­cle ex­hibits a de­struc­tive at­ti­tude as it sug­gests that in­ter­nal friend­ship and co­op­er­a­tion with con­ser­va­tive EAs is a waste of time, even though these re­la­tion­ships are vastly more pro­duc­tive than an or­di­nary friend­ship be­tween a Repub­li­can voter and a Demo­crat voter. Re­mem­ber that this is not part of the reg­u­lar Vox writ­ing—it is la­beled as part of the “Fu­ture Perfect” se­ries of EA jour­nal­ism.

Had Matthews ar­gued care­fully on a broader point that bi­par­ti­san friend­ship is in­effec­tive at im­prov­ing the poli­ti­cal sys­tem, this prob­lem would not ob­tain. Yet his im­me­di­ate leap to nar­row par­ti­san point of view, the poor qual­ity schol­ar­ship (even by op-ed stan­dards) and his epistemic un­fair­ness on the very topic of friend­ship with poli­ti­cal ri­vals com­bines un­char­i­ta­bil­ity with irony to make it strongly ex­clu­sion­ary. The ar­ti­cle heav­ily con­tra­dicts Ozy­man­dias’ recom­men­da­tions for wel­com­ing con­ser­va­tives in EA, al­though that was posted only yes­ter­day. And we should re­call that Matthews once wrote an ar­ti­cle that in­di­cated nega­tive at­ti­tudes to­wards the prevalence of whites in EA—it is rem­i­nis­cent of some of the at­ti­tudes sam­pled by Fowler et al and Schaf­fner et al. This looks like a dou­ble stan­dard. Of course, some peo­ple be­lieve that racial is­sues should be held to differ­ent stan­dards based on the his­tory of a par­tic­u­lar race. But even if you take that view, it should only ap­ply to in­sti­tu­tional ques­tions of whole racial com­mu­ni­ties which are in­trin­si­cally con­nected with racial his­tory. When we are work­ing with small tele­olog­i­cal com­mu­ni­ties and in­di­vi­d­ual peo­ple who have far more rele­vant char­ac­ter­is­tics be­yond their racial iden­tity, we need to re­mem­ber to treat them as such.

And lastly, what is Matthews try­ing to achieve with this ar­ti­cle? Are peo­ple spend­ing lots of time and money try­ing to make friends with poli­ti­cal en­e­mies just for the sake of fix­ing Amer­ica? The nom­i­nal ra­tio­nale for mak­ing this ar­gu­ment—that it will save young al­tru­ists from ex­pend­ing lots of time and effort on these civic du­ties—seems specious given how rare of a phe­nomenon that ac­tu­ally is.

The ar­ti­cle got min­i­mal at­ten­tion on so­cial me­dia, and I’m prob­a­bly mak­ing con­ser­va­tives feel worse about EA over­all by bring­ing it to their at­ten­tion. And it prob­a­bly seems weird that I’ve spent so much time and effort an­a­lyz­ing a rel­a­tively un­no­ticed op-ed from a jour­nal­ist. But I think it’s im­por­tant to set a clear early ex­am­ple for what sorts of be­hav­iors are con­sid­ered good or bad, and to build trust that some of us are mon­i­tor­ing this stuff and hold­ing it to se­ri­ous epistemic stan­dards.