Reductionist research strategies and their biases

I read an ex­tract of (Wim­satt 1980) [1] which in­cludes a list of com­mon bi­ases in re­duc­tion­ist re­search. I sup­pose most of us are re­duc­tion­ists most of the time, so these may be worth look­ing at.

This is not an at­tack on re­duc­tion­ism! If you think re­duc­tion­ism is too sa­cred for such treat­ment, you’ve got a big­ger prob­lem than any­thing on this list.

Here’s Wim­satt’s list, with some ad­di­tions from the parts of his 2007 book Re-en­g­ineer­ing Philos­o­phy for Limited Be­ings that I can see on Google books. His lists of­ten lack spe­cific ex­am­ples, so I came up with my own ex­am­ples and in­serted them in [brack­ets].

  1. Conceptualization

    1. De­scrip­tive Lo­cal­iza­tion: De­scribing a re­la­tional prop­erty as if it were monadic, or a lower-or­der re­la­tional prop­erty.

      • Fit­ness treated as a prop­erty of phe­no­type (or even of genes) rather than as a prop­erty of phe­no­type and en­vi­ron­ment.

      • [This may be equiv­a­lent to as­sum­ing that you can ap­ply lin­eariza­tion to re­move vari­ables from a func­tion. You of­ten do this to an­a­lyze the sta­bil­ity of equil­ibriums. So of­ten it’s a use­ful as­sump­tion.]

    2. Mean­ing Re­duc­tion­ism: As­sum­ing lower-level re­descrip­tions change mean­ings of sci­en­tific terms, while [go­ing back to?] higher-level re­descrip­tions [does] not.

      • Philoso­phers (who view them­selves as con­cerned with mean­ing re­la­tions) are in­clined to a re­duc­tion­ist bias.

      • [What he might mean: Modernist the­o­ries of mean­ing be­gin by study­ing sen­tences in iso­la­tion, as logic propo­si­tions. Take a clas­sic ex­am­ple, “Bach­e­lors are un­mar­ried men,” rep­re­sented as BAUM = forall(X, B(X)=>U(X),M(X)). They show that there are in­stances #bob where B(#bob) holds, yet af­firm­ing U(#bob) or M(#bob) would seem pe­cu­liar, and then say there­fore not(true(BAUM)), then gen­er­al­ize to say that true(forall(X, P(X)=>Q(X))) is not mean­ingful in gen­eral. But they don’t con­sider that we may use the English word “truth” differ­ently when talk­ing about propo­si­tions with quan­tified or un­bound vari­ables (BAUM) than when talk­ing about propo­si­tions with only bound vari­ables (“Justin Bie­ber is a bach­e­lor”). But I don’t think this is the main prob­lem with these mod­ernist analy­ses; com­pare the similar “birds can fly” /​ “*pen­guins can fly” ar­gu­ment.]

    3. In­ter­face Deter­minism: As­sum­ing that all that counts in an­a­lyz­ing the na­ture and be­hav­ior of a sys­tem is what comes or goes across the sys­tem-en­vi­ron­ment in­ter­face.

      1. black-box be­hav­iorism: all that mat­ters about a sys­tem is how it re­sponds to given in­puts

        • [Sys­tems with hys­tere­sis can­not be made sense of with a time­less black-box anal­y­sis.]

      2. black-world per­spec­ti­val­ism: all that mat­ters about the en­vi­ron­ment is what comes in across the sys­tem bound­aries and how it re­sponds to sys­tem in­puts.

        • [At first, this seemed true to me. But if the en­vi­ron­ment be­haves pre­dictably, and the sys­tem stud­ied re­lies on this pre­dictable be­hav­ior, any anal­y­sis that doesn’t model the en­vi­ron­ment will try fu­tily to find mechanisms in­side the sys­tem that pro­duce the needed in­for­ma­tion. For in­stance, if you were study­ing cir­ca­dian rhythms, and your model of the en­vi­ron­ment speci­fied the sky’s bright­ness at any par­tic­u­lar mo­ment, but didn’t model bright­ness as a 24-hour cy­cle, you would run into se­ri­ous difficulty try­ing to ex­plain the or­ganism’s cyclic be­hav­ior en­tirely in terms of in­ter­nal com­po­nents.]

    4. En­tifi­ca­tional an­chor­ing: As­sume that all de­scrip­tions and pro­cesses are to be referred to en­tities at a given level, which is par­tic­u­larly ro­bust, salient, or what­ever. This is the on­tolog­i­cal equiv­a­lent of as­sum­ing that there is a sin­gle cause for a phe­nomenon, or sin­gle level at which cau­sa­tion can act.

      • Thus the ten­dency to re­gard in­di­vi­d­ual or­ganisms as pri­mary [in se­lec­tion, pre­sum­ably].

      • Cf. method­olog­i­cal in­di­vi­d­u­al­ism for ra­tio­nal de­ci­sion the­o­rists and other so­cial sci­en­tists. [This is a par­tic­u­larly im­por­tant point to in­ject into the FAI/​CEV dis­cus­sion of “hu­man val­ues”. The val­ues en­coded into the be­hav­ior pat­terns of in­di­vi­d­ual hu­mans by in­di­vi­d­ual se­lec­tion, the val­ues en­coded by kin se­lec­tion, the goals they de­velop through in­ter­ac­tion with the en­vi­ron­ment (which are prob­a­bly not dis­t­in­guish­able, on later in­spec­tion of the brain, from “fi­nal goals”), the val­ues they hold con­sciously, and the so­cially-con­doned val­ues of hu­man groups, are all differ­ent, and en­coded in a va­ri­ety of rep­re­sen­ta­tions and lev­els of ab­strac­tions, and of­ten op­pose each other. A ra­tio­nal agent is by defi­ni­tion ra­tio­nal only within one rep­re­sen­ta­tion and with one set of non-con­tra­dic­tory goals. I haven’t seen dis­cus­sion in the FAI liter­a­ture of this prob­lem.]

      • Similarly for genes for some re­duc­tion­ist neo-Dar­wi­ni­ans. [Not sure if any­body ac­tu­ally holds such a po­si­tion.]

  2. Model Build­ing and The­ory Construction

    1. Model­ing Lo­cal­iza­tion: Look for an in­trasys­temic mechanism to ex­plain a sys­temic prop­erty, rather than an in­ter­sys­temic one. Struc­tural prop­er­ties are re­garded as more im­por­tant than func­tional ones, and mechanisms as more im­por­tant than con­text.

      • [I don’t know what he means by “func­tional”.]

      • [the ex­am­ple above of try­ing to model cir­ca­dian rhythms with­out mod­el­ing en­vi­ron­men­tal cy­cles would also be an ex­am­ple of this bias]

      • [Chom­sky posit­ing that chil­dren must have a built-in uni­ver­sal gram­mar be­cause he didn’t do the math]

      • [See all of be­hav­ior-based robotics and ev­ery­thing writ­ten by Rod­ney Brooks for ob­jec­tions to this bias in ar­tifi­cial in­tel­li­gence.]

    2. Sim­plifi­ca­tion: Sim­plify en­vi­ron­ment be­fore sim­plify­ing sys­tem. This strat­egy of­ten leg­is­lates higher-level sys­tems out of ex­is­tence or leaves no way of de­scribing sys­temic phe­nom­ena ap­pro­pri­ately.

    3. Gen­er­al­iza­tion: When start­ing out to im­prove a sim­ple model of sys­tem en­vi­ron­ment, fo­cus on gen­er­al­iz­ing or elab­o­rat­ing the in­ter­nal struc­ture at the cost of ig­nor­ing gen­er­al­iza­tions of elab­o­ra­tions of the ex­ter­nal struc­ture.

    4. Corol­lary: If the model doesn’t work, it must be be­cause of sim­plifi­ca­tions in de­scrip­tion of in­ter­nal struc­ture, not be­cause of sim­plified de­scrip­tions of ex­tre­nal struc­ture.

  3. Ob­ser­va­tion and Ex­per­i­men­tal De­sign

    1. Fo­cused Ob­ser­va­tion: Re­duc­tion­ists will tend not to mon­i­tor en­vi­ron­men­tal vari­ables, and thus will of­ten tend not to record data nec­es­sary to de­tect in­ter­ac­tional or larger-scale pat­terns.

      • [Nearly ev­ery drug tox­i­c­ity study ever, for failing to sam­ple each sub­ject’s gut micro­biome, which is a pri­mary de­ter­mi­nant of how in­gested drugs are bro­ken down]

    2. En­vi­ron­men­tal Con­trol: Re­duc­tion­ists will tend to keep en­vi­ron­men­tal vari­ables con­stant, and will thus of­ten miss de­pen­den­cies of sys­tem vari­ables on them. (“Ceterus paribus” is viewed as a qual­ifier on en­vi­ron­men­tal vari­ables.)

      • [Mouse ex­per­i­ments of­ten use seden­tary mice in HEPA-filtered cages en­vi­ron­ments fed ad-libi­tum. In­ter­ven­tions that ex­tend lifes­pan in such ex­per­i­ments, such as ra­pamycin or caloric re­stric­tion, may work less well in other en­vi­ron­ments.]

    3. Lo­cal­ity of Test­ing: Make sure that a the­ory works out only lo­cally (or only in the lab­o­ra­tory) rather than test­ing it in ap­pro­pri­ate nat­u­ral en­vi­ron­ments, or do­ing ap­pro­pri­ate ro­bust­ness analy­ses to sug­gest what are im­por­tant en­vi­ron­men­tal vari­ables and/​or pa­ram­e­ter ranges.

    4. Ab­strac­tive Reifi­ca­tion: Ob­serve or model only those things that are com­mon to all cases; don’t record in­di­vi­d­u­at­ing cir­cum­stances.

      • Raff (1996) [3] notes that evolu­tion­ary ge­net­i­cists fo­cus on in­traspe­cific vari­abil­ity, while de­vel­op­men­tal ge­net­i­cists fo­cus only on genes that are in­var­i­ant within the species. This pro­duces prob­lems both of method­ol­ogy and of fo­cus when try­ing to re­late micro-evolu­tion and macro-evolu­tion or evolu­tion and de­vel­op­ment.

      • Cog­ni­tive de­vel­op­men­tal psy­chol­o­gists tend to look only for in­var­i­ant fea­tures in cog­ni­tion, or ma­jor dys­func­tions, rather than pop­u­la­tional vari­a­tion.

    5. Ar­tic­u­la­tion-of-Parts (AP) Co­her­ence (Kauff­man/​Tay­lor/​Schank): As­sum­ing that stud­ies done with parts stud­ied un­der differ­ent con­di­tions are valid when put to­gether to give an ex­pla­na­tion of the whole.

      • [There’s a clas­sic case in cell biol­ogy of a chem­i­cal that has op­po­site effects on cells in vitro and in vivo, though I can’t re­call now what it is.]

    6. Be­hav­ioral Reg­u­lar­ity (Schank/​Wim­satt): The search for sys­tems whose be­hav­ior is rel­a­tively reg­u­lar and con­trol­lable will re­sult in se­lec­tion of sys­tems that may be un­char­ac­ter­is­ti­cally sta­ble be­cause they are in­sen­si­tive to en­vi­ron­men­tal vari­a­tions.

      • Schank: Reg­u­lar 4-day cy­clers among Sprague-Dawley rats are in­sen­si­tive to con­spe­cific pheromones. [This is prob­a­bly a refer­ence to this ar­ti­cle. I think Wim­satt’s point is that biol­o­gists chose to study ovu­la­tion cy­cles us­ing a par­tic­u­lar strain of rat be­cause it had reg­u­lar cy­cles, and it seems that that par­tic­u­lar strain of rat had reg­u­lar cy­cles be­cause of an in­bred ge­netic deficit in its reg­u­la­tion of ovu­la­tion cy­cles.]

      • [The ini­tial re­sis­tance to chaos the­ory and non­lin­ear sys­tems the­ory was due to lin­ear anal­y­sis hav­ing done a very good job for cen­turies on prob­lems that were stud­ied be­cause lin­ear anal­y­sis worked on them.]

  4. Func­tional Lo­cal­iza­tion Fallacies

    1. Deficit Reifi­ca­tion: As­sum­ing that the func­tion of a part is to pro­duce what­ever the sys­tem fails to do when that part is ab­sent, or pro­duced when it is ac­ti­vated or stim­u­lated.

      • spark plugs as “sput­ter sup­pres­sors”

    2. As­sum­ing 1-1 Map­pings Between Parts and Func­tions:

      • Stop­ping the search for func­tions of a part af­ter find­ing one; e.g., hemoglobin also func­tions in NO+ transport

      • Ig­nored di­vi­sion of la­bor when a part’s ne­ces­sity is shown through dele­tion stud­ies, thus miss­ing the roles of other parts

      • [The NCBI stores data on all bac­te­rial genes in a for­mat that as­sumes each gene has ex­actly one func­tion [4]]

      1. Ig­nor­ing in­ter­ven­tive effects and dam­age due to ex­per­i­men­tal ma­nipu­la­tions

        • in neu­ro­phys­iolog­i­cal studies

        • mark­ing spec­i­mens in mark-re­cap­ture stud­ies may af­fect their fitness

      2. Mis­tak­ing lower-level func­tions for higher-level ends, or misi­den­ti­fy­ing the sys­tem that is benefited:

        • [I think his ex­am­ples here are mis­taken]

        • elimi­na­tive re­duc­tion­ists who want to deny the ex­is­tence of large do­mains of cog­ni­tive func­tion [long dis­cus­sion of this here, which I recom­mend not read­ing; suffice it to say that I think the ERs be­ing com­plained of are not misi­den­ti­fy­ing the sys­tem benefited, but just us­ing lan­guage a lit­tle slop­pily]

        1. Im­po­si­tion of in­cor­rect set of func­tional cat­e­gories.

          • Com­mon in philos­o­phy of psy­chol­ogy when it ne­glects ethol­ogy, ecol­ogy, and evolu­tion­ary biol­ogy.

      Wim­satt (1980) then says:
      There are at least two pos­si­ble cor­rec­tive mea­sures… The first is ro­bust­ness anal­y­sis—a term and pro­ce­dure first sug­gested by Richard Lev­ins (1966) [2]. The sec­ond, which I will call “mul­ti­level re­duc­tion­ist anal­y­sis,” in­volves us­ing these heuris­tics si­mul­ta­neously at more than one level of or­ga­ni­za­tion—a pro­ce­dure that al­lows dis­cov­ery of er­rors and their cor­rec­tion....
      It should be clear that these heuris­tics are mu­tu­ally sup­port­ing, not only in their effec­tive use in struc­tur­ing and in solv­ing prob­lems, but also in re­in­forc­ing, in mul­ti­ply­ing, and, above all, in hid­ing the effects of their re­spec­tive bi­ases.… What­ever can be said for the­o­ries or paradigms as self-con­firm­ing en­tities, as much and per­haps more can be said similarly for [sets of] heuris­tics.

      [1]. William Wim­satt (1980). Re­duc­tion­ist re­search strate­gies and their bi­ases in the units of se­lec­tion con­tro­versy. In T. Nick­les, ed., Scien­tific Dis­cov­ery: Case Stud­ies, Dor­drecht: Rei­del, p. 213-259.

      [2]. R. Lev­ins (1966). The strat­egy of model build­ing in pop­u­la­tion biol­ogy. Amer­i­can Scien­tist, 54:421-431.

      [3]. Ru­dolf Raff (1996). The Shape of Life: Genes, Devel­op­ment, and the Evolu­tion of An­i­mal Form. Chicago: U of Chicago Press.

      [4]. They let you use mul­ti­ple GO tags, and put mul­ti­ple names within a pro­tein’s name field if sep­a­rated by slashes, but these are not ad­e­quate solu­tions.