Epistemology as a Social Science: Applying the Neuman-Rubin Method to Explain Expert Beliefs, in Carlo Martini & Marcel Boumans eds., Experts and Consensus in Social Science

Epistemology as a Social Science:

Applying the Neyman-Rubin model to Explain Expert Beliefs

Aviezer Tucker

I propose to naturalize parts of epistemology as social science.It is observable that some people have certain beliefs or atleast accept those beliefs as the best available. (Cohen 1992,Miller 2013, 1296-7) Standard surveys can establish correlationsbetween groups of people and clusters of beliefs that theyaccept. Hypotheses that explain the correlations can beepistemically significant and interesting. Such hypotheses maybe tested and corroborated using social science statisticalmethodologies that are more reliable than the methods of armchairepistemology like thought experiments.

I present here a social science approach to epistemicquestions about expertise such as: Which properties make peopleinto likely experts? What is the epistemic significance ofagreements among experts? Is there a special kind of expertagreement that is indicative of knowledge? Why do the beliefsexperts accept seem to enjoy a higher likelihood of beingknowledge than the opinions of non-experts?

The answers social science epistemology can offer to suchquestions are “external” in the sense that they do not trace theprocesses that generate knowledge, beliefs or their acceptance inthe minds of individual epistemic agents. Nor can social scienceepistemology examine the logical process of inference fromevidence. Social science epistemology is useful in cases when itis impossible or inconvenient to open and look into the mental ormethodological-inferential “black boxes” that connect cognitiveinputs with outputs, or when we want to look into those boxes butdo not know where they are. Non-experts may be unable to followcritically the process of belief formation of experts or may nothave the time and or other resources to do so. They may have toreach epistemic conclusions on the basis of social scienceanalysis of the group of experts. This resembles Goldman’sformulation of what he called “the novice-experts problem”: “somesorts of limiting factors-whether they be time, cost, ability, or

1

what have you-will keep our novices from becoming experts, atleast prior to the time by which they need to make theirjudgment. So the question is: Can novices, while remainingnovices, make justified judgments about the relative credibilityof rival experts? When and how is this possible?”1 (Goldman 2001,89)

Epistemologists interested in examining a type of inferenceof knowledge may not know where to start, where the expertisethat should generate knowledge is exactly, especially whenexpertise is contentious. In the case of, say, quantum physics,it is obvious that if there is any expertise in the field, it isheld by contemporary quantum physicists who publish papers in thefield in prestigious peer reviewed journals. However, in otherareas like economics, business, medicine, climatology, andenvironmental studies, the social location of expertise is not asobvious and may be contentious. Before epistemologists can studythe cognitive processes that generate epistemically significantoutputs, they need to know where to look, where expertise liesand who the experts whose practices they should study are likelyto be. For example, in attempting to explicate the epistemologyof our knowledge of the past, it is useful to know first whichgroup of experts is likely to possess such knowledge, who are thehistorians who are likely to possess knowledge of the past.(Tucker 2004)

The Neyman-Rubin model of causal inference is used byscientists, in particular social scientists, to infer types ofcauses of types of correlations. (Morgan & Winship 2007; Sekhon2010; Tucker 2012) Epistemology as a social science begins withepistemic hypotheses that connect types of epistemic causes (suchas knowledge, expertise, or types of bias) with correlationsbetween types of epistemic effects such as between types ofpeople and types of beliefs. The transition from tokens to typesof causal relations is achieved by the averaging of causaleffects2. For example, there is significant correlation betweenscientists and belief in Darwinian evolution, in comparison with1 I do not think that social science epistemology can distinguishbetween rival claims of experts. It can however distinguish likely from unlikely experts with rival claims.

2

the populations in which the scientists live. This does not meanthat each token scientist believes in a token of the theory ofevolution; only that on average, there is a higher correlationbetween scientists and belief in evolution than the correlationbetween comparable populations of non-scientists and their beliefin evolution. Another correlation may be between scientists andthe belief that the more public funds are allocated by thegovernment for scientists, the higher will be economic growth andjob creation. Again, not every token scientist must hold thistype of belief. On average scientists should have a greatertendency to accept it than comparable ordinary people. However,the cause of this correlation may be professional interest ratherthan expertise. The Neyman-Rubin method may help epistemologiststo distinguish between such causes of correlation betweenscientists and beliefs, between expertise and bias.

Generally, from an epistemic perspective, there are threetypes of causes of correlations between people and the beliefsthey accept: Expertise (which as I argue next means specialknowledge and impartiality), bias, and coincidence. The Neyman-Rubin is useful for finding out which of the three possible typesof causes is most likely to have been effective in causing thecorrelation.

2 It is possible to characterize a type as having a certainproperty even if not all its tokens have the property. Thoughmany tokens (for example of letters and words) resemble eachother and share properties, not all do: “The analogy to zoologyis helpful . Not every so-called black bear is black; not everygrizzly is four-legged, brown or has a hump . . . It may bepermissible to characterize the species in terms of suchproperties anyway. In many cases, one extrapolates fromproperties of the tokens, individually or collectively, toproperties of the type. However, . . . even if the overwhelmingmajority of the tokens have a property it does not entail thatthe type has it” (Wetzel 2009, 119, 120). Some properties of atype are not shared by any of its tokens; for example, “thegrizzly bear is endangered.” (Wetzel 2009, 120)

3

The Neyman-Rubin method infers types of causes of correlatedtypes of effects in two stages: In the first stage, it proves ordisproves that the correlations between the types of effects arelikelier given the hypothetical common cause type than givenalternative (unspecified) numerous types of causes. The testedcommon cause type hypothesis specifies the properties of the typeof common cause; but the properties of the alternative numeroustypes of causes are not specified3. For example, medicalresearchers are interested in finding out whether the correlationbetween an experimental group and healing is the result of thetype of treatment (medicine) all its members received or theresult of myriad other unspecified factors that affect healing.If both experimental and control populations are affected by thesame types of (unknown or unspecified) variables, but only theexperimental group is affected by a particular cause type (thetreatment), significant differences between the two populationsare likely to result from that cause type, e.g. the medicineincreased the correlation between the experimental group andhealing. If there is no significant difference in thecorrelations of the treatment and control groups with healing,the medicine probably was not causally effective.

Epistemologists, as social scientists using the Neyman-Rubinmodel of causal inference, may obtain a random sample of apopulation and divide it into two randomly assigned groups, whoseonly difference is the presence or absence of an epistemicallyinteresting type of cause like a cognitive input or a biasingfactor that the experimental group is introduced to but thecontrol group is not. It is possible then to measure the averageeffect of the input or bias on the cognitive outputs (knowledge,beliefs, acceptance of beliefs) of the experimental group bymeasuring if there is a significant epistemic output difference

3 I emphasize that the inference is of a common cause type ratherthan common cause token. Confusions between the inferences ofcommon cause types and tokens have been rife and destructive inphilosophy since Reichenbach. I have argued that the distinctionbetween the inference of common cause types and tokensdistinguished the theoretical from the historical sciences.(Tucker 2007; 2012)

4

between the beliefs of the two groups. Such experiments arecommon in behavioral economics and have produced most interestingresults.

At the second stage, assuming that the correlation betweentypes of effects is more likely given the specified type of causetype than given numerous unspecified different types of causes,the Neyman-Rubin method attempts to find the exact causalrelations or nets, which may be complex, requiring theconstruction of multicollinear, interactive, and so on models. Inthe process, the method needs to eliminate possible confounders,common cause types of both the hypothesized type of cause and thecorrelation between its apparent effects. For example,personalized medical attention may cause both the administrationof medicine and healing. Types of stress may cause both smokingand cancer. If statistically significant number of members ofthe experimental group are cured except a sub-group that ishomogeneous in say, sharing the stage of development of thedisease or in having a particular genetic makeup, it is stillwarranted to assign the cure to the new drug because the bestexplanation of the exceptional, dissenting, group is itshomogeneous makeup.

In the experimental sciences, experimental designs controltypes of variables to isolate their effects. When suchexperiments are impossible, scientists resort to statisticalobservational data analysis. Using a variety of statistical controltechniques to hold different variables constant while measuringothers, social scientists conduct multivariate regressionanalyses that generate causal maps that measure levels of causalinfluence that each variable exerts on the others.

“Natural,” undesigned, experiments, can be just as effectivein confirming hypothesis as laboratory experiments when naturecooperates. The experimental group should be as identical intype to the control group as possible. For example, in studyingthe causes of a type of cancer, random assignment is impossible.Instead, scientists select a random sample of people with cancerand a random sample of the population that contains the cancersufferers and compare them to see if some causes seem to be more

5

significantly correlated with the first than with the secondgroup, e.g. smoking.

Expertise as the explanation of belief

In the following discussion I consider expertise to meanimpartiality and special knowledge. Goldman (2001) discussed themeaning of expertise at some length. I agree with him in broadterms about the link between knowledge and expertise, but I addimpartiality as a necessary condition. Knowledge could notgenerate experts unless they are also impartial; otherwise, theiraccepted beliefs would reflect their biases rather theirknowledge.

“Anthropological” reliance on professional organizations oracademic institutions, affiliations and certifications toidentify experts is a tempting alternative pragmatic approachthat we use in everyday life to navigate society. At their best,professional and academic institutions should promote,promulgate, regulate and enforce expertise that makes agreementsamong their members epistemically significant and likely.However, historically, this appeal to authority has failed toooften. Academic and professional institutions have been toosusceptible to political pressure and coercion from without,economic interests and graft from within, internal power andseniority hierarchies, or the shared biases of a class of peoplewho share professional interest and sometimes social background.The alternative approach I endorse does not have to rely oninstitutional authority.

The question is whether expertise in the sense ofimpartiality and special knowledge is a better explanation ofcorrelations between groups of people and clusters of acceptedbeliefs than coincidence or bias. The Neyman-Rubin method cantest the expertise hypotheses first against the coincidencehypothesis and if it passes that test, against the biashypothesis.

People may adopt beliefs or even reach consensus on them(perfect correlation) coincidentally, just like an experimentalgroup in medicine can be cured irrespective of the medicine

6

because of the various different properties of the patients; or agroup of people with lung cancer may have become sickirrespective of their smoking because they were exposed to myriadother risk factors. A complex alternative theory to thehypothesis that expertise accounts for a significant correlationbetween a group and its accepted beliefs, may combineconjunctions and disjunctions of different factors to explain thecorrelation. Miriam Solomon (2001) advocated such explanation ofscientific consensus. Solomon suggested that scientific consensuson beliefs results from many different “decision vectors” thatsomehow have the same effect. Better scientific theories areaffected by a greater variety and distribution of decisionvectors than theories that follow a more narrow range of them, epluribus unum. Solomon did not regard ‘decision vectors’ such asideology or pride, deference to authority or agreement withscriptures, as impediments to the achievement of scientificknowledge, but as inevitable necessary prerequisites forscientific progress. If Solomon is right, experts such asscientists do not agree because they share expertise, despitetheir different genders, cultures, interests and so on, butexactly because they are men and women of different cultures andof various interests and so on. Solomon collapsed the distinctionbetween evidence (‘empirical vectors’), cognitive values(‘theoretical vectors’), and biases (‘social vectors’).Altogether, Solomon (2001, 62–3) estimated there are 50–100decision vectors of all kinds that affect theory choice bythemselves or in interaction with each other4.4 Solomon suggested that diversity of decision vectors is thebest explanation of consensuses on beliefs among scientists.Still, ceteris paribus, even without using the Neyman-Rubin method,the likelihood of consensus is higher given a single common causetype such as expertise or an overwhelming bias for all theopinions of the experts than given many diverse decision vectorsbecause of the low probability that all different biases willcause identical types of beliefs. In Bayesian terms, it wouldrequire multiplying the independent likelihoods of the beliefsgiven each independent type of bias by each other. Even if eachlikelihood is high to begin with, the result of multiplying manysuch factors by each other would reduce the total likelihood to

7

Miller (2013, 1307-8) introduced the historical example ofbroad acceptance by most medical scientists that excess acidityin the stomach caused peptic ulcers. In Miller’s opinion, thisconsensus resulted from several decision vectors: Academichierarchical status and power disparity between powerfulsupporters and weak deniers of the acidity hypothesis; prevailingphysiological theory that held that the stomach was too acidicfor bacteria to survive in; and the economic interests ofpharmaceutical companies in a chronic illness that required alifelong supply of anti-acids rather than a single round ofantibiotics. Miller acknowledged that there were competinginternal explanations of this episode in the history of sciencethat did not resort to social biases. It does not matter for thecurrent discussion if Miller’s historical example is correct ornot. Miller is right that “on occasion” coincidences can and dohappen and various interests, biases, and so on coalesce in highcorrelation with the same set of beliefs. The same can happen inmedicine, despite proper experimental design, significantstatistical gap between treated and control groups can followdifferent causes rather than the treatment. Still, ceteris paribusin most cases the treatment is more likely to be the cause of thedifference than many different causes. This conclusion isprobable and not certain and hence fallible. The Neyman-Rubinmethod acknowledges that outcomes such as beliefs can be affectedby the coincidence of large number of known and unknown factors(or vectors). But if a control group (general non-expertpopulation) is very likely to share those vectors with theexperimental group (of experts), they probably cancel each otherand expertise remains last standing as the best explanation ofthe doxastic outcomes.

The Limits of the Neyman-Rubin method in epistemology

close to zero. Miller (2013) noted correctly that this criticismis valid only when the vectors should explain a consensus in thesense of absolute agreement among all. A weaker but stillsignificant correlation may be likely given different causes thatgenerate the correlation with high likelihood.

8

The Neyman-Rubin method can generate interesting epistemicresults. But it can function only within bounds: Statisticalsamples of the tested “expert” group must be large enough tocancel out the various vectors of the larger control, non-expert,group. Smaller samples would not necessarily reproduce thevarious “decision vectors” or biases of the control group and mayhave a higher concentration of some decision vectors and so maygenerate results (beliefs) that do not result from the expertise“treatment.” For example, a consensus on an esoteric topic by ahandful of professionals may follow peculiar social dynamics incomparison with society, for example the domination of a singleor few authoritative figures, “a guru” or “a cabal” (Goldman2001, 98) that determine who is and is not hired in a smallacademic social milieu, and consequently who teaches andpublishes. Such correlation or even consensus cannot function asan indicator of expertise according to the Neyman-Rubin methodbecause the sample is too small, it does not replicate thedistribution of properties in the population at large, and it hasa higher concentration of authoritarian control by a small cabalthan in society at large where many different concentrations ofpower can countervail each other5. This can happen only inisolated academic oligopolistic hierarchical structures where asmall cabal can leverage its control over other institutions

5 Goldman’s (2001) analyses the position of a layperson decidingbetween experts in terms borrowed from the epistemology oftestimony. Only independent testimonies matter and they shouldbe evaluated according to their reliabilities. Experts havehigher reliabilities than lay people about esoteric subjects theyspecialize in. Despite the Bayesian framework, Goldman left outthe prior estimation of the probability of what the expertstestify for. As Bovens & Hartmann (2003) demonstrated, when thatprior is low enough, two even unreliable but independenttestimonies are sufficient for inferring knowledge. ThisBayesian epistemology of testimony approach cannot discriminatebetween competing expert opinions if the experts who testify areindependent and have similar credibility and their testimonies-opinions are similarly surprising.

9

lower down on the hierarchy to impose its opinions. Onceexpertise goes global and outside the academy, it spins out ofcontrol and no oligopoly can usually achieve control any morethan any financial cartel can control global markets and pricesfor a substantial period of time in today’s global economy. Forthis reason, the correlation between scientists and Evolutionaryand Relativity theories or historians and much of what they agreeon about the past (which conspiracy theorists dispute) cannot bereduced to a small sample with a rigid social hierarchy, as insome parts of the academic humanities.

The Neyman-Rubin method is also useless for identifyingexpertise if there are multiple groups with inconsistent beliefsthat nevertheless seem to have the same expertise and the samesocial composition in comparison with the same control group ofnon-experts. The Neyman-Rubin method is useless for choosingwhich group is most impartial and knowledgeable. The uniquenessof the social correlation between the group of people affected byexpertise and beliefs is essential for utilizing the Neyman-Rubinmethod in epistemology. If more than one group possesses thesame expertise but forms different conclusions, an externalsocial science method cannot identify which one is more likely tobe in possession of impartial knowledge.

Multiple sub-groups with strong correlations withinconsistent beliefs are not a problem for the Neyman-Rubinmethod if all the groups of people who have inconsistent beliefsto those of the “expert” group are homogenous and their beliefscan be explained by that homogeneity. In such a case it can beargued that the otherwise heterogeneous expert group is the onlyimpartial group. For example, the claim for impartial expertiseof the scientists who believe in Darwinian Evolution is supportedby the fact that all those who deny evolution, belong to fairlyhomogenous religious groups, while believers in evolution includemembers of virtually all religions. Similarly, there arehistoriographic accounts that are accepted by all or almost allhistorians except homogenous partisan groups. For example, theHolocaust, the deliberate genocide of about six million Jewsduring the Second World War, is believed by all the experts onthe Second World War irrespective of their national, religious or

10

ideological background. All Holocaust deniers are Neo-Nazis ofvarious shades.

Confounding the Confounders

When it is seems likely that a correlation between a group andits beliefs is the result of expertise and not numerous otherunknown factors, epistemic social scientists must be wary ofconfounding factors, either alternative single causes toexpertise that also distinguish the expert from the controlgroup, or common causes that affect both the expertise and thecorrelation between the group of experts and its acceptedbeliefs. If not the result of coincidence, the significantcorrelation between a group and its accepted beliefs may resultfrom bias.

An alternative common cause type to expertise can becoercion. Coercion is no foundation for real consensus, but for‘unwilling acquiescence’. (Caws 1991, 379) Coercion may take manyforms, violence, intimidation, threats, manipulation by economicdependence, or plain browbeating. Even oppressive coercion israrely sufficient for coercing a full consensus of expressedopinions. Some people are not easily intimidated, others have astrong character and do not react to threats, and others taketheir own beliefs sufficiently seriously to express themregardless of the effect it may have on their personal fortunes.Historical experience demonstrated that the establishment of evena local coerced consensus on public expressions of belief in astate-sanctioned dogma required the extensive and extreme use ofviolence. Both the Inquisition and the Soviet NKVD secret policecould not coerce without killing dissenters. Still, coercion canaffect the correlation between groups of people and the opinionsthey express, whether or not they believe in them. Coercion maylead to the establishment of a discursive hegemony. Coercivebodies, most notably the state, may be interested only in thepublicly accepted opinions of experts known for their specialknowledge and expertise. This interest rather than expertise maycreate the apparent significant doxastic gap between the expertand control, non-expert, groups; the state may not care for thepublic opinions of the uneducated masses.

11

Confounders that affect both the expertise and thecorrelation between the group of experts and their beliefs mayinclude social factors that select the experts and then affecttheir judgment. For example, if all medical doctors were maleand they considered the male body the standard human body, thereis a strong case for claiming that their beliefs about humananatomy as well as the social criteria that selected them tobecome experts reflected male biases rather than expertise.Masculinity confounded expertise.

If social science epistemologists eliminate knownconfounders, there is a fallible case for expertise. For example,scientists who endorse Darwinian evolution are an incrediblyheterogeneous group. It is difficult to imagine one or moreconfounding biasing factors that would turn the apparentlyheterogeneous scientific community into a homogeneously biasedgroup in a way that would confound the hypothesis that expertisecauses the significant correlation between scientists and beliefin Darwinian evolution. Yet, this conclusion is fallible. It isalways possible that social science epistemologists overlooked adoxastic biasing factor, and an apparently heterogeneous samplewas actually homogenous in a doxastic relevant way. A newhypothesis that discovers a confounding bias in a group mayundermine its claim for expertise. For example, feministphilosophers of science developed hypotheses that discoveredcorrelations between male biases in scientific communities inmedicine and biology and their scientific beliefs (cf. Longino1990, 103–214; Okruhlik 1994). Before the introduction offeminist philosophy of science, historians and philosophers ofscience surely noticed that all doctors were males, but did notconsider the gender composition of scientific groups significantfor the explanation of their beliefs. Before the hypotheses, thegroup of medical researchers was considered to be composed ofexperts.

Conversely, a hypothesized confounding bias may be rejectedby using the Neyman-Rubin method. For example, Communistideologues claimed that Mendelian genetics and Darwinianevolution were biased by capitalist free market competitionideology. The Nazis claimed that the concentration of scientists

12

of Jewish descent in branches of sciences like relativity theorybiased them as “Jewish science.” But the social background ofgeneticists and the religious practices of the ancestors ofphysicists were irrelevant for the acceptance of their knowledgeas impartial and for recognizing them as experts. From a socialscience epistemic perspective, all expertise must emergesomewhere, sometime, among some people. The question, whenapplying the Neyman-Rubin method to epistemology, is whether overtime the beliefs that emerged and developed among a specific,even homogenous, group of experts become accepted by verydifferent kinds of people who then serve to refute social biasconfounding hypotheses and support an expertise hypothesis thatexplains better the high correlation between people of verydifferent backgrounds and accepted beliefs, in Mendelian geneticsand relativity theory for example.

In natural experiments, especially because of the absence ofrandom assignment, it is possible to guard only againstconceivable confounding biases that are connected withalternative hypotheses to the expertise hypothesis. It is alwayspossible that a relevant confounding bias has been overlooked andthe expertise hypothesis is false despite the apparentheterogeneity of the consensus group. Confidence in the expertisehypothesis is directly related to the variety and diversity ofalternative confounding hypotheses. The more varied alternativeconfounding hypotheses there are, the easier it is to guardagainst bias.

Cognitive values

One particular possible confounder may undermine the whole ideaof implementing in epistemology the Neyman-Rubin method fordiscovering expertise: shared cognitive values. Arguably,expertise is just an intermediate variable between sharedcognitive values and correlations with beliefs. If so, expertagreement is not a reflection of knowledge and impartiality, butof shared cognitive values.

Cognitive values determine which statements are worthy ofbeing considered knowledge, absolutely or in comparison withcompeting statements. Arguably, cognitive values may confoundexpertise. Kuhn (1996, 184–6) suggested that the scientific

13

community is constituted by cognitive values: accuracy,consistency, scope, simplicity and fruitfulness. Cognitive valueschanged more slowly than theories in the history of science. Ifshared cognitive values are necessary for forming beliefs, theexpertise hypothesis may have to be at the very least qualifiedas relative to particular sets of cognitive values as biases thatare shared by the putative expert group.

It is possible to reapply the Neyman-Rubin method toepistemology to examine cognitive values. The conducivenesshypothesis explains the high correlation between experts and typesof cognitive values in comparison with a comparable control groupof non-experts by their greater conduciveness to knowledge, incomparison with other cognitive values. For example, the birthof science was accompanied by a shift from cognitive values thatconsidered knowledge to be the result of faith, revelation,ancient wisdom and tradition, to empiricist values that arearguably more conducive to knowledge. The conducivenesshypothesis can then be supported by the Neyman-Rubin method justlike the expertise hypothesis, and be fallibly corroborated.Experts who possess impartial special knowledge unsurprisinglycorrelated more strongly with cognitive values that are conduciveto the attainment of knowledge than comparable groups that arenot in the business of obtaining special knowledge, or fail toachieve it. Competing hypotheses that claim to explain asignificant correlation of experts with cognitive values by anyother (particularly external social or cultural) variables wouldfind it quite difficult to explain the appeal of these values tovery different experts and their lower appeal to very similargroups of non-experts. The size of the group (not a sample) thatupholds scientific cognitive values is much larger than that ofany group of experts with special knowledge; it is the set of allthe sets of experts. Accordingly, a comparable control group ofnon-expert will have to be just as universal. The heterogeneityand sizes of these groups should prove any bias in theachievement of expert correlation with scientific cognitivevalues highly unlikely.

Comparison with Consensus accounts of Expertise

14

The Neyman-Rubin method’s fallible inference of expertise fromsignificant correlations between sufficiently large, uniquelyheterogeneous, and unconfounded groups and accepted beliefs, incomparison with control groups is preferable to alternativeepistemic attempts to use consensus or consensus inducingprocedures as markers of knowledge, rationality, and expertise.The application of the Neyman-Rubin method to social scienceepistemology should make epistemology clearer, more parsimonious,pragmatically applicable to everyday contexts rather than toideal situations, better reasoned, and politically unproblematic.

A literally universal consensus about any belief in anygroup is very rare, a utopian ideal with few instantiations inthe real world. (cf. Miller 2013, 1297) Epistemologists whoresorted to a concept of consensus had to specify a particularmeaning of this concept that could make it useful or resort tovagueness: ‘Consensus is typically not all encompassing… sinceusually some dissent remains. The decision whether or not to calla state of affairs “consensus” or “dissent” is to some extentarbitrary.’ (Solomon 2001, 118 my italics A.T.)

One popular approach substitutes a process or a procedurethat should generate a consensus for the consensus itself. Forexample, in Rescher's (1993) interpretation, Habermas limited hisphilosophic discussion of consensus to an ideal rational type,according to an ideal process. But an account of epistemic consensusthat reduces it to a process that should generate the proper,significant, type of consensus begs the question. If a consensusis epistemically significant only if it follows a particularprocess, then the process is epistemically important, not theensuing consensus. Similarly, the use of consensus to elucidaterationality must presuppose an account of rationality to discoverwhich consensus is rational. Nicholas Rescher (1993, 13)criticized what he took to be Habermas's concept of consensus asbegging the question of rationality. Rescher noted that if‘consensus’ is used in the sense of an ideal consensus under idealcircumstances that allow for a process of unconstrained rationalconsideration, consensus is reduced to the outcome of unbridledrationality. Unbridled rationality does not require a consensusor even a process. If a single individual or a machine possessesall the relevant evidence, background knowledge and pure

15

rationality, a rational result will be inferred inevitably and aconsensus would be superfluous.

An epistemic focus on an ideal consensus excuses ignoringthe absence of actual literal consensuses where all experts agreeon all their expert beliefs as well as actual cases of consensusthat were irrational or founded on false beliefs, as deviationsfrom the ideal. Habermas followed Peirce in attempting tounderstand truth by analyzing an ideal consensus that shouldfollow a reflexive process of argumentation, mutual criticism,and airing of cultural values6. Since for Habermas only theresults of an ideal speech situation were rational, less thanideal speech situations, practically all real situations ofagreement and disagreement in society, were to varying degrees,less than rational. It would stand to reason then for experts toproceed on what an ideal speech situation would have resulted in,rather than bother with necessarily imperfect actualizations ofthe ideal of consensus. (Habermas 1990, 67)

Since for Habermas only the results of an ideal speechsituation are rational, less than ideal speech situations areless than rational. It is rational for a Habermasian to proceedon what an ideal speech situation would have resulted in, ratherthan bother with a necessarily imperfect actualization insomething like anarchic democratic process. An intellectualminority can then conveniently define and represent the

6 Charles Peirce (1877, 132-133) believed that thescientific process will ultimately generate consensus onobjective truth. His faith in an unspecified scientific processof investigation was combined with an eschatological rhetoricabout consensus as the ‘destiny’, ‘destined center’,‘foreshadowed Goal’, and ‘predestined opinion’ of the history ofscience. Consensus for Peirce was not concrete, but a chiliasticideal to be realized at the end of time, a form of epistemicmessianism. Before the promised consensus comes about at the endof the scientific process, it is impossible to know which beliefsare true and which will be forsaken during the future history ofscience, just as it is impossible to distinguish the righteousfrom the sinners, the city of god from the city of man, beforejudgment day.

16

hypothetical rational consensus that would have resulted from anideal procedure. Historically, this was the Jacobininterpretation of Rousseau’s general will, a hypothetical generalwill that could not be ascertained empirically.

The Jacobin legitimized their dictatorship by claiming thatthe opinions of the people do not count, are not the generalwill, as long as they are hungry, dominated and misled by therich who hire venal journalists to manipulate them (Talmon 1970,106). Jacobin “democracy” is not about listening to the peopleor counting their votes, but about the creation of conditions forthe true expression of the general will. Till then, theintellectual revolutionary vanguard may follow the common will,not democratic majority vote (Ibid, 105-107). “The majority inthe real sense is where the true general will resides, even ifthat will happens to be expressed by a numerical minority” (Ibid,99) of Jacobin activists. The idea that popular majority votedoes not count as democratic until certain egalitarianprerequisites are satisfied runs from the Jacobins through toHabermas. Habermas’ utopian vision, like that of the Jacobins,was the homogenization of society, an end to power distinctionsand power biased “knowledge.” Accordingly, the second task ofthe rational intellectual or expert who know which consensusshould emerge from a rational process of deliberation is toobtain the kind of political power that would allow them todestroy all competing power centers to create the homogenoussociety that can then finally reach a rational consensus, theJacobins reign of terror.

“Only in an egalitarian public of citizens thathas emerged from the confines of class and thrown offthe millennia-old shackles of social stratification andexploitation can the potential of an unleashed culturalpluralism fully develop—a potential that no doubtabounds just as much in conflicts as in meaning-generating forms of life. But in a secularizedsociety that has learned to deal with its complexityconsciously and deliberately, the communicative masteryof these conflicts constitutes the sole source ofsolidarity among strangers….” (Habermas 1996, 308)

17

Rescher (1993, 156-157) traced back Habermas’ utopian consensusto Hegel’s march of reason in history and Marx’s dream of ahomogenous society without special interests. The totalitarianpower required for such radical egalitarian transformation ofsociety, to allow the emergence of “authentic completedemocracy,” has never been used for that purpose. Once any groupachieved totalitarian power, it used it in its own interests.The Jacobins first excluded from discussion the specialinterests, the upper classes and the clerics; when that provedinsufficient, they decapitated them, when that proved stillinsufficient they proceeded to decapitate other classes andspecial interests, culminating with themselves.

Other theories within the Habermasian orbit proposeddifferent and sometimes less radical lists of necessaryconditions for the emergence of epistemically significantconsensus. Helen Longino (1994) argued that the consensus thatcounts must result from critical dialogues among individuals andgroups from different points of view, an ‘interactive dialogiccommunity.’ (1994, 142–3) Longino stipulated further theproperties of the community that can transmute subjective pointsof view into scientific objectivity: there must be public forafor criticism; the community must be responsive to criticism;there must be revisable public standards, cognitive values, towhich participants in the discourse can appeal to reachagreement; and the community should be intellectuallyegalitarian, where decisions about beliefs are not taken byappeal to authority. Critical discourse should explain howdifferent biases and points of view are transformed intoconsensus on beliefs. Still, like Habermas, Longino could not beinterested in consensus per se, but in an ideal process forgenerating knowledge. The conditions she stipulated have beensatisfied very rarely. These conditions are also too narrowbecause there can be a consensus or broad agreement among expertswithout deliberation, when a conclusion is obvious, givencognitive input, and discussion is superfluous, for example, whena surprising piece of evidence is discovered or the results of acrucial experiment become known.

The use of the Neyman-Rubin method releases epistemologyfrom having to rely on vaguely defined or ideal consensus or from

18

being forced to allow ideal presuppositions and procedures tosubstitute for the elusive consensus. From the perspective ofthe Neyman-Rubin model, a consensus is just an extreme andhistorically rare case of statistically significant correlationbetween a group and beliefs. Statistically significantcorrelation in comparison with a control group does not stipulateany particular epistemic process, nor does it demand thesatisfaction of utopian preconditions whose self-defeatingsatisfaction requires political dictatorship. Epistemologistswho are interested in the procedures that produce the significantcorrelation with beliefs can then pursue a different, intrinsic,research project inquiring how the group of experts reached itsbeliefs.

Boaz Miller (2013) proposed that knowledge is likely to bethe best explanation of epistemic consensus the more threeconditions are met: Apparent consilience of evidence, sharedevidential standards in a consensus community, and diversity ofsocial backgrounds. Miller’s three conditions are intrinsic (therelation between types of evidence and knowledge), pragmatic(evidential standards and use of language that describes theevidence), and social (diversity of backgrounds). However, ifthe first criterion is well satisfied, the other two areredundant. If epistemologists can trace and judge the relationbetween evidence and beliefs, measure the consilience of theevidence and how it supports the posterior probability ofhypotheses, the other criteria do not matter for the epistemicevaluation of whether the support of the evidence to thehypothesis is sufficient for considering it knowledge. Forexample, suppose the epistemologist concludes from studying therelation of evidence to hypothesis that it is well corroborated,though all the scientists who accept the hypothesis are whitemale zombies. Would the possible zombie bias matter here?!Arguably, not, because the epistemologist can trace intrinsicallythe process of inference from evidence irrespective of theidentity of the experts. The “zombie hypothesis” that sharedbeliefs may reflect zombie culture and biases is relevant only ifwe cannot, or do not have the resources, to check the evidentialjustification of the hypothesis ourselves. Miller defended hissocial diversity requirement against this sort of argument,

19

claiming that without sufficient social diversity the evidencecannot be sufficiently consilient. Social diversity shouldcontrol against two kinds of biasing influences, expected andunexpected. Known hypotheses connect expected biases such as thefinancial interests of researchers with the results of theresearch. “[D]iversity is also required for controlling forunexpected influences. There may be influences of which we areunaware, and the more diversified the consensus, the less likelythey will affect it. This is similar to the rationale behind thedesign of randomized clinical trials, which are required tocontrol for both known confounders (sex, age, etc.) and unknownconfounders.” (Miller 2013, 1312) Miller implies that theNeyman-Rubin method can guard against both expected andunexpected biases and he is right; this is what I have beenarguing for in this paper.

Diversity is not the only way to guard against biases. Itshould be possible for qualified observers with sufficientresources to detect faulty research design or inference fromevidence irrespective of expected or unexpected biases.Diversity or the Neyman-Rubin method is useful when it isimpossible or too expensive in terms of time and other resources toexamine the relation between evidence and accepted beliefs, orwhen we want to trace this process but are not sure where to lookfor it. In this second case, the Neyman-Rubin method may besocial scaffolding that holds an epistemic edifice while it isbeing built.

Down with the Scaffoldings!

Epistemology as a social science has the advantage of attemptingto build an edifice of expertise “from the ground up,” from theobservable social effects of epistemic processes, rather than“top down,” from ideal definitions of knowledge and rationality.

Epistemology can use the results of the search for expertiseto locate a social and historical space of expertise.

Sufficiently large and uniquely heterogeneous groups thatare strongly correlated with sets of beliefs, more thancomparable control groups, enjoy a fallible presumption ofexpertise in the absence of confounders. Members of such groups

20

are likely to be experts. This presumption of expertiseprecedes any knowledge and analysis of their epistemic practices.Epistemologists can use then the Neyman-Rubin method to knowwhere to look for expert special knowledge. Epistemologists whoare interested in an internal epistemic account of the generationof knowledge by experts, can investigate further whichmethodologies generate expert knowledge within each putativegroup of experts. I applied this method in the philosophy ofhistoriography (Tucker 2004). Before attempting to answerquestions about the epistemology of our knowledge of the past, Iasked first where, when and among which group of experthistorians such knowledge has emerged, spread, and becomeaccepted. The answer to this question provided a natural basisfor an epistemology of the knowledge of the past. This external,social scientific conclusion allowed me to build an epistemologyof historiography from the ground up, by studying the theoriesand methodologies that enabled the emergence of this uniquelyheterogeneous group that reached such a broad agreement about somuch of what it believed about the past.

Similar epistemic applications of the Neyman-Rubin methodcan benefit other epistemic sub-fields that attempt to understandknowledge derived from expertise, like the philosophies of thespecial sciences. Epistemic inquiry can then investigate whatkind of special knowledge expert groups possess that allow themto reach significant correlations with beliefs. This correspondswith standard scientific search for mechanisms, afterestablishing correlations between the treatment and thecorrelated results, for example, if the medicine correlates withhealing in comparison with a control group, or if smokingcorrelates with higher rates of cancer among smokers incomparison with a non-smoking comparable group, scientists askfor the physiological and chemical mechanisms that cause thehealing or the cancer. Similarly, epistemologists can find outhow particular groups of experts obtain their knowledge, afterthey establish the likelihood of their expertise, and drop thescaffolding of social science epistemology.

Bibliography

21

Bovens Luc & Hartmann Stephan (2003) Bayesian Epistemology. Oxford:Oxford University Press.

Caws Peter (1991) ‘Committees and Consensus: How Many Heads AreBetter than One.’ Journal of Medicine and Philosophy 16 375 91

Goldman, Alvin (2001) “Experts: Which Ones Should You Trust?”Philosophy and Phenomenological Research, 63, 85-110.

Habermas, Jürgen (1990) Moral Consciousness and Communicative Action,trans. Christian Lenhardt & Shierry Weber Nicholson, CambridgeMA: MIT Press.

Habermas, Jürgen (1996) Between Facts and Norms: Contributions to aDiscourse Theory of Law and Democracy, trans. William Rehg, CambridgeMA: MIT Press.

Kuhn Thomas S (1996) The Structure of Scientific Revolutions, 3rd ed.Chicago: Chicago University Press

Longino, Helen E. (1990) Science as Social Knowledge: Values and Objectivity inScientific Inquiry, Princeton: Princeton University Press.

Longino, Helen E. (1994) "The Fate of Knowledge in SocialTheories of Science," in Socializing Epistemology: The Social Dimensions ofKnowledge, Frederick F. Schmitt ed., Lanham MD: Rowman &Littlefield, 135-157.

Miller, Boaz (2013) “When is consensus knowledge based?Distinguishing shared knowledge from mere agreement,” Synthese,190, 1293-1316.

Morgan, Stephen L. and Winship, Christopher (2007) Counterfactualsand Causal Inference: Methods and Principles for Social Research. Cambridge:Cambridge University Press.Okruhlik, Kathleen (1994) “Biology and Society,” Canadian Journal ofPhilosophy, Supplementary Vol. 20, 21-42.

Peirce Charles S 1958 [1877] ‘How to Make Our Ideas Clear.’ inSelected Writings. Philip P. Weiner ed. New York Dover 113-136.

Rescher Nicholas (1993) Pluralism: Against the Demand for Consensus.Oxford The Clarendon Press

22

Sekhon, Jasjeet S. (2010) “The Neyman-Rubin Model of CausalInference and Estimation cia Matching Methods.” In The OxfordHandbook of Political Methodology, Janet M. Box-Steffensmeier, Henry E.Brady, and David Collier, eds., Oxford: Oxford University Press,271–299.Solomon Miriam (2001) Social Empiricism. Cambridge MA MIT Press .

Talmon J. L. (1970) The Origins of Totalitarian Democracy, New York:Norton.

Tucker, Aviezer (2004) Our Knowledge of the Past: A Philosophy ofHistoriography. Cambridge: Cambridge University Press.

Tucker, Aviezer (2007) “The Inference of Common CauseNaturalized,” in Jon Williamson & Federica Russo eds., Causalityand Probability in the Sciences. London: College Press, 439-466

Tucker, Aviezer (2012) “Sciences of Tokens and Types: The Difference between History and the Social Sciences,” in Harold Kincaid ed., The Oxford Handbook of Philosophy of the Social Sciences. Oxford:Oxford University Press, 274-297.

Wetzel, Linda (2009) Types and Tokens: On Abstract Objects. Cambridge,MA: MIT Press.

23