First published in What if there were no significance tests? (Eds.) Harlow, Lisa L., Mulaik, Stanley A., and Steiger, James H. 1997. New Jersey: Erlbaum. Extract. pp. 366–391.1 Good Science is Abductive, Not Hypothetico-Deductive Abduction A century ago, Charles Peirce (1839-1914) introduced the term abduction to de- note a type of practical inference that he viewed—rightly—as importantly distinct 2 fromclassicallyconceiveddeduction(logicalentailment)andstatisticalinduction. This is our creation of explanations for observed phenomena, a mode of reasoning that not merely expands upon the information at hand (“ampliative” inference), as statistical generalization to relative frequencies in populations also does, but hypothesizes why these observations have their distinctive character. And these explanations invoke concepts other than the ones used to describe the observa- tions explained. Thus from the Hartshorne and Weiss (1934) collection of Peirce’s papers: Abduction consists in studying facts and devising a theory to explain them. (1903; Vol. 5, p. 90). Abduction is the process of forming an exploratory hypothesis. (1903; Vol. 5, p. 106). Abduction must cover all the operations by which theories and con- ceptions are engendered. (1903, Vol. 5, p. 414). And most provocatively, The great difference between induction and hypothesis [that is, abduc- tion] is, that the former infers the existence of phenomena, such as we have observed in cases which are similar, while hypothesis supposes somethingofadifferentkindfromwhichwehavedirectlyobserved,and 1(Ed.) In the earlier part of the paper, not included here, WR discussed the Hypothetico- Deductive and Bayesian approaches to scientific inference. He now turns to the third option, Explanatory Induction, which he refers to here using C. S. Peirce’s term “abduction.” 2Peirce attributed this term to a Latin translation of Aristotle’s Prior Analytica, lib. 2, cap. 25 (Moore, 1984, p. 108). Peirce also refers to abduction as reasoning by “hypothesis” and, less often, “retroduction.” 1 frequently something which it would be impossible for us to observe directly. (1857; Vol. 2, p. 388.) Peirce returned repeatedly to this theme without, however, developing details on how to do it. And he urged that abduced hypotheses be appraised by follow- up tests of their additional implications insomuch as they generally remain far 3 from a sure thing given just the observations that originally invoke them. . So it might be argued that Peircian abduction was but a softer, more tolerant precursor of Popper’s fiercely astringent hypothetico-deductivism, which insisted both that there is no such thing as a “logic of discovery” (i.e., that hypothesis creation is purely a psychological phenomenon governed by no principles of rationality) and that hypotheses never gain credibility save through verification of their deductive consequences. But unlike Popper, Peirce would have welcomed the prospect that some abductions have determinate forms which transmit conviction. I shall argue that such inferences are the machinery of knowledge acquisition in both technical science and everyday life. Be that as it may, the notion of “abduction” remained quietly where Peirce left it for the first half of this century during which our discipline’s method ortho- doxies were codified. But the massive mid-century swing of philosophic zeitgeist rejecting positivistic epistemology included an awakening of philosophers’ desire to acknowledge abduction somehow. Hanson’s Patterns of Discovery (1958) was warmly received as something of a breakout despite its arguing merely that our observations are impregnated with theory from the outset while endorsing a view of abduction as the intuitive onset of not wholly implausible conjectures, without concern for any identifiable principles that might govern these. And philosophers have also become increasingly inclined to speak sagely of “inference to the best explanation” since Harman (1965) introduced this phrase as a new take on Peir- cian abduction, even though neither Harman nor, until quite recently, anyone else had much to say about which hypotheses are explanatory or what conjectured explanations qualify as decent much less best. However, artificial intelligence (AI) work on problem solving expressly identified as Abduction (see especially Josephson & Josephson, 1994) has begun to put algorithmic muscle into the “best explanation” slogan. Meanwhile, Simon (1973) argued in explicit opposition to Popper that there does indeed exist a logic of scientific discovery, and has since developed programmable details within his own AI framework for problem solving (Langley, Simon, Bradshaw, & Zytkow, 1987). And in Rozeboom (1961), follow- ing my personal epiphany on deriving theory from data during graduate work on 3Thusinthe1903lectures: “[Abduction’s]onlyjustificationisthatfromitssuggestiondeduc- tioncandrawapredictionwhichcanbetestedbyinduction.”(Hartshorne&Weiss,1934,Vol.5, p. 106) 2 4 the behaviorist What is learned? problem, I pointed out specific logical forms by which discerned data regularities are intuitively transformed into explanations thereof. Originally, I called this “ontological induction” because it creates con- cepts of entities (attributes and events) distinct from the ones described by the data statements. But later, in a user-friendlier and more mature statement of the position (Rozeboom, 1972), I renamed it “explanatory induction” to stake out the 5 seminal role I claim for it in real-life knowledge acquisition. Although differing considerably in detail, all these revisionist views of scientific inference fit nicely within the broad tent of Peircian abduction. Even so, the version I have called explanatory induction has a tightly operational focus warranting recognition as a distinct species within Peirce’s genus. ‘Abduction’ is at risk of becoming a buzzword in AI circles; and the extent to which psychology’s research methods can profit from study of AI data-processing algorithms claimed to be abductive is problematic. A salient case in point is the sector of AI problem solving covered by Josephson and Josephson (1994), which has a considerable history of papers with ‘abduc’ in their titles. This has focused on deriving from a plurality of observations on a single individual— notably, med- ical symptoms and lab workups in the studies published, but also potentially data clusters such as features of a crime scene, or of a handwriting sample, or of style in artwork of uncertain authorship, etc.—the best diagnosis of that individual’s con- figurationofunderlyingconditionsresponsibleforthosesymptoms. Buttheextant AI algorithms that accomplish this are preprogrammed with “domain knowledge” containing all the explanatory if/then laws posited to produce symptoms of the sort to be interpreted in particular cases. Algorithms that can do this effectively in difficult cases may well have applied value as expert systems; but they tell us nothingabouthowtoconceiveandacquireconfidenceintheexplanatorylawsthey presume. (Elsewhere in the AI literature, programs can be found that profess to abduce laws as well; but the versions I have seen use a production logic more suited to fantasizing than to scientific inquiry.) In contrast, Simon and colleagues have said nothing in print about “abduction” or “inference to best explanation”; but their induction modeling has attempted to reconstruct discovery of laws and concepts that were historically important achievements in physical science. These 4Whatcameasrevelationtomewasrealizationthatalthoughnon-mentalisticS-Rmediation mechanismscouldinprincipleaccountforacertainprospectivetransfer-of-trainingphenomenon thatcommonsensewouldtaketomanifestamentalistic“idea”mediatingbetweenexternalstimuli and behavior, this phenomenon would demand explanation by a certain structure of mediation with indifference to whether that structure has a mentalistic, S-R mechanistic, or some other embodiment. (See Rozeboom, 1970, pp. 120–122). 5MyoriginallabelprovestobethesuperiorversioninthatInowwanttoemphasizethatsta- tisticalinduction,too,vergesuponexplanationwhenitreachesbeyondpopulationfrequenciesto underlyingprobabilities. However,Ialsosubmitthatprobabilities,thoughundeniablytheoretical entities, are not themselves explanatory mechanisms but only supervene upon those. 3 prestigiousscientificsuccessesareparadigmexamplesoftheinferencesIhavecalled explanatoryinduction; andIwelcomeLangleyetal’s(1987)progressinsimulating these as support for my still-minority thesis that interpreting observations in this fashion is the engine of advance in epistemically praiseworthy scientific belief. Explanatory induction, or “EI” for short (not to be confused with the “AI” of artificial intelligence), has two special epistemic merits typical neither of abduc- tion in Peirce’s broadly inclusive sense nor of output from the artificial intelligence programs that have been labeled abductive by their creators: Not merely do ex- planatory inductions call explanations for observations to mind in the first place bydecentlydeterminateinferenceforms, theyalsoyieldconfidenceintheirconclu- sionsthatinfavorablecircumstancescanapproachthevividstrengthofperceptual beliefs. So why has this essay’s title paid explicit homage generically to abduction ratherthanspecificallytoEIwhenthelatterisitsrealfocus? Threereasons: First, good science does indeed profit from broadly abductive (imaginative) speculations so long as results from innovative research provoked by these are not interpreted by blue-sky HD reasoning. Second, ‘abduction’ is a word that for better or worse is bound to become increasingly prominent in discussions of scientific method; so you may as well get familiar with it. And third, it is clumsy style to use an eight-syllable phrase in slogans or titles when three syllables will do. Explanatory Induction (EI): Examples Supposeyourhandcalculator—callitcpfor“thiscalculatoratpresent”—isacting strangely: Whenever you enter numeral 8, its rightmost display cell responds with 6, and the same occurs when the result of a calculation should show 8 in that dis- play position. Your conclusion is immediate and assured: Something is wrong with CP. Were you to observe further that cp responds to entry of any digit other than 5, 6, or 8 with a non-numeric shape in its rightmost display cell never containing an upper-right vertical stroke, you would probably conclude more specifically that the upper-right pixel at end of cp’s display isn’t working. But even if no digit other than 8 manifests a problem in this display position and you know nothing about cp’s display mechanism, you are still confident that some feature of cp has changed for the worse even though all you yet know about this altered condition is that it makes cp show 6 in its rightmost display cell under input circumstances that formerly would have put ‘8’ there. In either case, you have committed an act of explanatory induction on these observations from which you would have been incapable of abstaining. You have inferred that cp is in some state that degrades the desirable input/output relations it sustained previously. And you do not take thisstatetobeanephemeralmanifestpropertythatcphasjustatmomentswhen depression of a key co-occurs with a display inappropriate to that key, but which vanishes when no key is pressed or when key entry and consequent display are 4 in agreement (e.g., when entry and right-most display digit are both 5). Rather, you feel sure that it is an enduring property, one that is responsible for some key presses eliciting incorrect responses and persists in cp even when this is resting or giving correct responses. What you know about a properly working cp is even more richly endowed by EI. First, even if you had never received instruction in how to use cp beyond its on/off switch and advice that pressing cp’s keys (which we will describe as “entries” corresponding to the keys’ assorted labels) generally alters cp’s display, youcouldstillhavebecomeconfident,throughobservationsoncp’sdisplaychanges following key entries, of many generalities having form (1) Whenever cp is in condition C and then receives sequence K of key entries, its display at end of this input sequence is R, wherein C specifies, among other observable preconditions, cp’s display at start of entry sequence K and some information about cp’s recently prior entries. In practice (and you really have acquired such beliefs about hand calculators), these generalities would for the most part accrete silently in your “background knowl- edge” while entering your conscious judgment as particularized anticipations of how cp’s display should respond to a possible key-entry sequence initiated here andnow. Andthesalientpointtotakefromthisiswhatyoutherebybelieveabout cp at various stages in a sequence of its key entries. To become aware of beliefs that in your normal use of cp are too fleetingly transient to reach foreground attention, suppose that your execution of some cal- culation with cp is interrupted by a phone call. On return, you observe that cp’s current display is 5 and, instead of hitting cp’s clear-all key to start afresh, you wonder what its display would become were you to enter a few more digits, say 9, 3, 9, followed by a function key such as the one labeled ‘=’. You realize that you can’t predict cp’s future display just from the present one and your choice of input starting now, but can do so if you remember enough of your transaction with cp just prior to breaking for the call. And if you are fairly sure that your last function-key entry was × while cp’s display then was 1.2, you can anticipate with some confidence that were sequence 939= now to be entered, cp’s display imme- diately following the = entry would be the product of numbers 1.2 and 5939 or, more precisely, a digital representation of that which after some paper-and-pencil scratching you can identify in advance as digit string 7126.8. But your success in predicting cp’s displays in this situation isn’t the point here. What matters are your beliefs (a) that cp’s display R at times t′ later than the present t will be ′ lawfully determined in part by its sequence of key entries between t and t; (b) that the present state of cp at t also makes a large difference for R at t′; and (c) that although this state of cp so crucial to its subsequent display is ephemeral 5 and not directly observable, its functional character—that is, its distinctive role in production of cp’s behavior—can be inferred from past input to cp and described by the conditional (2) If the symbol string that starts with cp’s current display excluding terminal dotandthereafterdescribesthenextsequenceK of cp’skeyentriesrepresents a number r, and entry of K is immediately followed by one of function keys +,−,×, ÷,or =, then cp’s resulting display will be standard notation for the number that equals 1.2-times-r, to which in afterthought you append some auxiliary if-clauses constraining the size of r and how roughly cp is handled. Idiom urges that (2) be simplified to (3) cp is currently disposed to respond to any input number with the product of that with 1.2, and may even tempt you to say, more metaphorically than EI approves, that cp remembers 1.2 and is in the mood to multiply. Metaphor aside, the idiom of (3) makes it easy for you to acknowledge that when cp is working properly, it is capable of passing through many transitory states, each describable by a covert conditional—“covert” in that it is an if/then only implicitly—of form (4) cp is disposed at time t to respond to input of any number r with s⊙r, wheres isarealnumberand⊙isoneofbinaryoperators+, -, ×,÷. Sinceatmost one completion of schema (4) is true of cp at any particular time t, (4) in effect identifiesatwo-dimensionalarrayofattributealternativeswhichwemightcallcp’s “op(eration)-states.” From there, your theory of how the s and ⊙ facets of cp’s op-state at any given time have been brought about by input is a straightforward inference from your generalizations of form (1). In resurrection of the positivist program early this century for clarifying sci- ence’s theoretical concepts, one might argue that cp’s op-state at any time t is nothing morethan somelogicalconstruction out of cp’s keyentriespriortot. But your intuitive EI leap from observations on cp to beliefs such as (4) about cp’s op-states insists that its property described by (4) is a contemporary event which, though due to and predictable from input/output events in cp’s past, is distinct from those and mediates whatever causal influence those may have on cp’s next 6 response. Of course, this interpretation of (4) could be wrong. But for better or worse it is the conclusion that EI delivers here. 6To be sure, if you had the flu on this date last year, then it is a property of you-today that you had the flu a year ago. But we view such temporal displacements of natural events as linguistically contrived epiphenomena that supervene on the world’s causal unfolding. (If you 6 The op-states of cp inferred by EI in this example are atypical of EI discovery both in their extreme ephemerality and in the accuracy with which cp’s op-state at any given time can be predicted from its prior input history. But they usefully demonstrate the compulsive power with which explanatory inductions can occur. Given enough observations on cp to convince you that the form-(1) generaliza- tions you have taken from these will continue to fit future observations on cp, it is psychologically impossible for you not to advance from there to indicative and counterfactual conditionals like (2) that in turn are arguably equivalent in mean- ing, or nearly so, to dispositional claims of form (3). These moves are far from epistemicallyunproblematic; butinsomuchastheyareasnaturalasbreathingand nearly as indispensable, prudence advises us not to disown them but to develop expertise in their technical management. I have taken pains to describe cp’s op-states as “dispositions” because that is the established label for an enormous variety of qualities ascribed to things in everydaylife: thesournessoflemonsvs. sweetnessofsugar, youruncle’sstinginess vs. the generosity of your aunt, the fragility of chalk vs. the durability of wood, the stickiness of honey and adhesives vs. the slipperiness of teflon and wet ice, and so on for thousands of adjectives in ordinary language. Whatever may be the nature of what we attribute to the individuals said to be that way, it is deeply ingrained in our language that these are dispositions, or “tendencies” if you prefer; and over many decades ofphilosophers’ efforts to reduce their auraof mystery (see Tuomela, 1977, for a collection of modern views) it has been more or less agreed that these are essentially the same as what we attribute by claims of form (5) If S(x), then (probably) R(x), or variations and elaborations (e.g. conjunctions) thereof. In this, ‘x’ is place- holder for names of whatever entities we may wish to characterize this way, ‘S(x)’ describes an input condition that can be imposed on x, and ‘R(x)’ describes a response of x that may, though need not, be some display by another object to which x is coupled in some fashion specified in ‘S(x)’. (That is, R(x) can be a meter reading.) Also, the conditionality expressed in (5) is understood to tolerate occasionalfailureofresponseR giveninputS.Thatis, whateverwemeanby(5)in its disposition-demarking sense, not only does truth of ‘R(x)’ not suffice for (5) to becorrect,itisalsoallowedthat(5)mightholdinsomeinstancesevenwhen‘S(x)’ is true while ‘R(x)’ is false. In particular, when x belongs to a set X of entities that we think are alike in respect to (5)—notably, when X comprises a succession havepersistinghealthproblemstoday,thismaywellbedueinparttoyourboutoffluayearago throughtheiterateddynamicsofcertainbodychangesinitiatedatthattime;butitissurelynot brought about by your having today the property of having had flu a year ago.) This may only exhibit the naivete of our causality intuitions; but it’s still the way to go until the error of those becomes more apparent. 7 of an enduring thing’s temporal stages over some limited time interval, or stages of different things that appear to be interchangeably similar in many ways—we are disposed to accept (5) if, among the members of X whose S and R conditions have been observed, R is considerably more prevalent among Xs that have been Sdthanamongthosethathavenot. Indeed,insuchcaseswefindourselvestalking about different degrees of the If-S-then-R disposition (which is thereby converted from an inferred state to a dimension of inferred states) such that the strength of this common to the members of an X presumably homogeneous in this respect is measured by the contrast of R’s relative frequency among X-members who have been observed in condition S to its relative frequency observed among X-members lacking S. Ordinary-language versions of such dispositional concepts usually specify their input/output conditions so vaguely and so categorically (that is, treating S and R as sloppy condition-present/condition-absent dichotomies) as to be nearly useless for scientific research save as targets of replacement by more precisely defined and more finely graded input/output alternatives. Sometimes such improvements can be viewed as more accurate diagnoses of roughly the same underlying attribute di- mensions detected crudely by their ordinary-language precursors. But more often, invention of new measuring instruments—devices or special environments whose carefully controlled coupling with objects that interest us scarcely ever occurs naturally—afford display of precisely defined dimensions of quantifiably graded response alternatives that initially provide conception, and thereafter finely dis- criminatingdetection,ofourstudiedobjects’underlyingpropertiesbeyondtheken of everyday experience. These previously hidden properties have now become vir- tuallyobservableevenifnotquitesomanifestastheinput/outputconditionsfrom which we infer them. And data on patterns of co-occurrence among these newly discernible attributes may—or may not—lead us through iterated explanatory in- ductions to understanding of even deeper levels of inner mechanisms responsible for these subjects’ overt behavior. Psychology has had its full share of instrument-grounded disposition concepts, though apart from physiological psychology the “instruments” have largely been special stimulus settings with constraints on subjects’ movements therein rather than the meters, gauges, chemical reagents, and other sensor devices familiar in everyday life through spillover from the physical and biological sciences. Indeed, much of educational psychology and personality assessment has been grounded on subjects’reactionstonumerousbrieftestitems—paradigmatically,choiceamonga small number of alternative answers to a written or sometimes spoken question— that are collected into “scales” on which the subject receives a numerical score summarizing responses to the items comprising that scale. Much has been written and still more remains to be said about the technology of such questionnaires, especially about the rationale of item grouping which in practice often illustrates 8 EI at work on a deeper level with techniques of factor analysis. But we cannot address that here. The point to be taken is that educational and psychological measurements of this sort are classic illustrations of explanatory-inductive con- cept formation accompanied by collection of data on the variables so inferred, even though their attendant circumstances seldom warrant trust at the higher confidence levels of which EI is capable, such as when these same subjects’ body weight is measured by a balance or torsion scale, or their temperature by a ther- mometer. Each item’s definition (standardized manner of stimulation and method of scoring) creates conception of a mini-tendency to get one score rather than another if so stimulated. High EI confidence that mini-dispositions so conceived genuinely exist requires subjects to show consistent differences in their responses to each item over repeated testings, which in practice is seldom demonstrated firmly. But there also exist other patterns of manifest intrasubject consistency in response to a battery of test items that urge inference to relatively enduring “traits” of tested subjects. These are measured, though with less-than-perfect ac- curacy, by subjects’ composite scores over selected subsets of the test items and are inferred to dispose not merely subject-distinctive patterns of response to these particulartestitemsbutalso—whichmayormaynotbelaterconfirmed—toother input conditions as well. (See McCrae & Costa, 1995.) In practice, we seldom have much interest in dispositions identified by just one dimension of response to just one specific stimulus setting. But EI kicks in hard when we have observed a cluster of dispositions expressible in ideally simple cases by a plurality of sentences having form ‘When a thing of sort B is disposed to R when S d’ then almost always it is also disposed to R when S d.’ i i j j (The S envisioned here are a considerable diversity of conditions that can be i imposed on B-things, each R , is a response made possible by input condition S , i i and B is some background condition—often conceived only vaguely by reference to a few paradigm examples—under which this consilience of mini-dispositions seems dependable.) In such cases, EI waives the mini-dispositions in favor of a single theoretical property τ whose presence/absence in a thing of kind B can be diagnosed in diverse known ways (whether the thing does R in response to i S for several different S -tests) and is moreover expected to partake in yet-to- i i be-discovered lawful relations with other observable and EI-inferable variables as well. Actually, finer details of these S /R tests (notably, when some of the R are i i i graded response alternatives) usually yield conception of this τ as a theoretical variable taking a considerable range of alternative values over things of kind B. Such “cluster concepts” (as some philosophers have called them) of classic simplicity abound in chemistry, mineralogy, and medicine wherein the “natural kind”ofachemicalormineralor,inmedicine,thepresence/absenceofaparticular diseaseconditionisdiagnosedbyabatteryofsuchtests. Apowerfulcaseinpointis the chemical contrast originally conceived as acid vs. alkali vs. salt. Partington’s 9 (1935/1965)historyofchemistrylistsdozensofdifferenttestsdescribedbyRobert Boyle (a 17th Century founder of modern chemistry) wherein a sample x of some to-be-appraised material X is brought into contact with a sample s of some test material S . For suitably chosen S , changes in the appearance of s resulting from i i this contact reliablyforecasts, for many other identified materials S what changes j in samples of S will result from their contact with samples of material X. (Many j though by no means all of these test outcomes are changes in s’s color that depend in part on the chosen S . Another especially important response in some tests is j s’s dissolution in liquid x.) Early chemists did not learn about acids and alkalis by first speculating that such theoretical properties might exist, next deducing observableconsequencesofthishypothesis,andfinallyconfirmingthosepredictions asatriumphofhypothetico-deductivescience. Rather,thealkali/salt/acidnotions and their eventual refinement into a continuum of pH levels were an explanatory induction from observed patterns of reaction such as collated by Boyle. Or so I submit. Another good example from everyday life is your learning in childhood about hot. When in the kitchen, or near a fireplace, or perhaps in a family workshop, you had repeated experiences wherein touching a certain object resulted in your immediatelyfeelingacutediscomfortfollowedalittlelater,ifyourtouchwasfirmor prolonged, by blistering of your skin at the point of contact. You easily convinced yourselfthatanobjectwhichsoaffectedyouwhentouchedwoulddosoeverytime youtoucheditforashortdurationthereafter,thoughusuallynotaftersomelonger lapse of time. So you concluded—not by deliberated reasoning but by wired-in cognitive compulsion—that some things at certain times have an If-I-touch-it-I’ll- get-hurt property. Andadditionalfoolingaroundorwatchingothersdealwithsuch objects also taught you that a thing having this touching-it-hurts-me property is also featured by If it is touched by a plastic object the plastic will melt, and by If fat is spilled on it the fat will sizzle and likely catch fire, and by If a scrap of paper is held against it the paper will turn curly brown and maybe burn, and by If a pan of water is put on it and it stays able-to-hurt-me long enough, the water will boil. Indeed, youlearnedthatanyoneoftheseif/thensholdingforanit-now prettywell guaranteed that the others were true of it-now as well, whence you concluded— again by innate urge though not quite so compulsively as before—that all these simultaneous if/thens probably manifest a single underlying condition that you came to think of as “hot” because you also learned that a nearby grown-up’s shouting ‘hot’ when you were reaching for something also dependably indicated thatyourtargetwasinthisdangerstate. Fromthere, youwentontodiscoverthat a thing’s glowing red often signaled that it was hot, that devices your folks called “thermometers”canfinelydiscriminatedifferencesinhot thatyoucouldpreviously distinguishonlycoarsely,andsoonforalargerepertoireofbeliefsaboutgradations of hot so securely rooted in your direct observations that even today you may have 10