Thursday, January 30, 2014

Falsification

The topic is getting some play in the blogosphere due in part to the post on this matter by Sean Carroll (the physicist at Cal Tech, not the eve-devo biologist) on the Edge question of the year (see here and find his contribution on page 202-3). At any rate, there has been some interesting discussion from people in other areas that you might be interested in looking at (here, here, and here). The discussions give subtle takes on Popper's falsificationist dicta, all recognizing that the (at least common) understanding of Popper is way off. Pigliucci notes the problems that the Duhem-Quine thesis raise, i.e. the observation that theories are never tested directly but always rely on many auxiliary assumptions that can serve to insulate a thesis from direct empirical refutation. The discussion is interesting and may serve to temper simple minded resort to the falsificationist gambit.

One thing, however, that seems not to have been mentioned is that quite often we can find evidence for claims that are not falsifiable: for example, the germ theory of disease postulated germs as agents of disease transmission. It's not clear that existential statements are falsifiable, viz. the absence of attested black swans does not imply that black swans don't exist. We make many such existential claims and then go out and look for instances, e.g. germs, black holes, Higg's particles, locality conditions, etc.  In other words, we often actively look for confirmatory evidence for our claims rather than looking to refute them. In fact, I might go further, in the early exploratory phase, finding evidence for one's views is actually more important that finding evidence that could refute them. Good ideas start out life as very fragile. There is always tons of apparent evidence that "refutes" them. Indeed, the more daring the proposal, the more likely it appears to be false (and perhaps the more likely that it is false). So, what does one do? Look for places where it works! And this is a smart thing to do.  It is always fair to ask someone why their proposal matters. One way that it can matter is that it does something interesting, i.e. solves a puzzle, explains recalcitrant data, predicts a new kind of object etc.  Evidence in favor of a proposal is what allows one to make demands on a rational listener. So, falsification as a strategy is useful, but only after a proposal has gained admittance as a live possibility and admission to this august group is paid for in verifications.

At any rate, the discussion is pretty good and there are links to other matters that might matter. It's always nice to see how others flounder with these large methodological concerns. Enjoy.

20 comments:

  1. Carroll pretty clearly just smuggles falsifiability back in under his assertion that good theories are empirical.

    "It's in that sense that the success or failure of the idea is ultimately empirical: its virtue is not that it's a neat idea or fulfills some nebulous principle of reasoning, it's that it helps us account for the data."

    If a theory has empirical consequences, there's a possibility that those consequences will disagree with reality. In Carroll's words, if a theory cannot account for data, this might be because the theory is, in fact, false. Given Duhem-Quine, this is, ultimately, (a fairly subtle form of) falsifiability.

    ReplyDelete
    Replies
    1. Pretty much everyone agrees that it's good for scientific theories to have empirical content in some sense. Popperian falsificationism is a much more specific doctrine than that for a couple of reasons. First, we would standardly take existential statements such as "there are black swans" to have empirical content, but such statements are not strictly falsifiable. Second, Popper has a particular conception of scientific methodology according to which falsification (and crucially not anything called "induction") plays a central role. I think Carroll clearly has quite a different conception.

      Delete
    2. My point is just that, whatever Carroll means by empirical, it includes a notion of falsifiability. Perhaps it's a more subtle form of falsifiability than Popper put forth, but he's not just defending the empirical content of existential statements. He wants theories that "account for the data," and that opens the door to refutation and falsification and the like.

      To the extent that Carroll is arguing against anyone who insists that falsification constitutes the full set of properties that define science (or good scientific theories), he's arguing against an idea that few, if any, people (other than maybe Popper) have championed and that has long been understood to be, at best, a small part of the full picture.

      Delete
    3. Well, he's arguing against a Popperian conception of the scientific method. No doubt he would accept that theories should be falsifiable in the very broad sense in which you're using the term. I don't think that means that he's "smuggling falsifiability back in", since he never rejected falsifiability in your broad sense to begin with.

      Delete
    4. This comment has been removed by the author.

      Delete
    5. >>To the extent that Carroll is arguing against anyone who insists that falsification constitutes the full set of properties that define science (or good scientific theories), he's arguing against an idea that few, if any, people (other than maybe Popper).

      It is simply a misunderstanding to think that Popper was *just* a falsificationist. He was quite happy with confirmation. In fact, his whole attempt was to try and understand under what circumstances confirmation was meaningful. For him, confirmation of a theory was meaningful only if the theory was falsifiable (however, he seems to have "uniqueness" in mind when he used the word "falsifiable", as I discuss below).

      In the reference below you will see that Popper was as much a confirmationist as anyone else. In fact, his view appears to be very similar to the Bayesian perspective on confirmation (which is quite something to say given the historical disagreements). You will also see that though he uses the word "falsification", he seems to have meant uniqueness (re: "in fact with results which everybody before Einstein would have expected").

      Now, it is possible that Popper was inconsistent in discussing his views in other works; I am not an expert on his stuff, so I don't know. But, I don't expect anyone to be perfectly consistent. And, what he says seems perfectly reasonable.

      So, one could paraphrase a part of what Popper said (at least this incarnation of him) as saying that a confirmation of a theory is impressive if it involves the confirmation of a prediction that is unique to the theory (or different from other relevant theories). If phrased this way, it seems to be extremely palatable, and not very much at odds with what Norbert says above.

      [Ref: http://www.stephenjaygould.org/ctrl/popper_falsification.html]

      Delete
    6. Thanks for the details about Popper. I wrote "other than maybe Popper" because I was assuming that even Popper wasn't all and only about falsification.

      Well, he's arguing against a Popperian conception of the scientific method.... I don't think that means that he's "smuggling falsifiability back in", since he never rejected falsifiability in your broad sense to begin with.

      Based on Karthik's comment, it sounds like he's not actually arguing against what Popper proposed. I probably should have just pointed out that he's arguing against a straw man, instead of arguing that he's smuggling falsifiability back in.

      Delete
    7. Maybe a straw man in Philo circles. I think that Carroll notes this in passing, observing that falsification is no longer the methodology of choice of the sophisticated. However, if a scientist wants a cudgel to beat a disliked view, then falsification comes easily to hand. I think that this is the view that C wants banished. The piece by Pigliucci makes the point that C, in defense of multiverses and strings may be letting himself off the hook too easily in suggesting that falsification is the problem the critics suffer from. If so, then here indeed is a case of attacking a straw man even given his own criteria. At any rate, nice to see that naive falsification is now taken to be an obvious non starter.

      Delete
    8. @Karthik, @Noah. I find it odd to read Popper as a confirmationist. Popper was quite explicit that our best scientific theories are very likely to be false. I would think a confirmationist is someone who thinks just the opposite: that our best scientific theories are (or at least should ideally be) highly probable. From the Stanford Encyclopedia article:

      In the view of many social scientists, the more probable a theory is, the better it is, and if we have to choose between two theories which are equally strong in terms of their explanatory power, and differ only in that one is probable and the other is improbable, then we should choose the former. Popper rejects this. Science, or to be precise, the working scientist, is interested, in Popper's view, in theories with a high informative content, because such theories possess a high predictive power and are consequently highly testable. But if this is true, Popper argues, then, paradoxical as it may sound, the more improbable a theory is the better it is scientifically, because the probability and informative content of a theory vary inversely—the higher the informative content of a theory the lower will be its probability, for the more information a statement contains, the greater will be the number of ways in which it may turn out to be false. Thus the statements which are of special interest to the scientist are those with a high informative content and (consequentially) a low probability, which nevertheless come close to the truth. Informative content, which is in inverse proportion to probability, is in direct proportion to testability. Consequently the severity of the test to which a theory can be subjected, and by means of which it is falsified or corroborated, is all-important.

      These views led Popper to develop the slightly hokey notion of "verisimilitude", which was supposed to play the role standardly assigned to "truth" or "probable truth" (no small role to play!)

      I think that because of Popper's huge influence on pop scientific culture, a lot of people underestimate how wacky his views actually were. Carroll's critique probably wouldn't worry a sophisticated neo-Popperian (if there are any), but I'd say it's fairly squarely aimed at The Logic of Scientific Discovery.

      Delete
    9. @Alex: I think we are using different definitions of "confirmation". I am using the one mentioned in Ref1 below ("The term “confirmation” is used in epistemology and the philosophy of science whenever observational data and evidence “speak in favor of” or support scientific theories and everyday hypotheses").

      To discuss confirmation in terms of probabilities of hypotheses is forcing a Bayesian interpretation on hypotheses - and this is not a standard definition as far as I know.

      In the sense of confirmation that I defined above, Popper was clearly a confirmationist, as can be seen from a careful reading of [Ref2].

      I don't like second-hand discussions of ideas because they often mis-state the original position. As far as I can see, that is no less true of the reference you provided. Crucially,

      "But if this is true, Popper argues, then, paradoxical as it may sound, the more improbable a theory is the better it is scientifically, because the probability and informative content of a theory vary inversely—the higher the informative content of a theory the lower will be its probability, for the more information a statement contains, the greater will be the number of ways in which it may turn out to be false."

      As you can see from Ref2, this does not follow from what Popper actually said, at least in the few things that I have read. It is simply a misunderstanding of Popper that is carried forward, especially by philosophers of science of the Bayesian persuasion. Popper argues that confirmation of a more falsifiable theory is more impressive (more accurately "confirmation of a more unique prediction", if uniqueness has degrees :)).

      It is only in a Bayesian view that a more falsifiable theory is less probable. Hypotheses/Theories having probabilities is a Bayesian view, and does not hold true for other viewpoints [As a tangent: Deborah Mayo has spilt a lot of ink on why probabilities over hypotheses/theories might be problematic]. Which is exactly why some Bayesians have misunderstood Popper - the error lies in reinterpreting Popper in Bayesian terms. Therefore, it is also an error to reason from what Popper wrote that he suggests a less probable theory is better than a more probable theory when both are equally explanatory.

      [Ref1: http://www.iep.utm.edu/conf-ind/]
      [Ref2: http://www.stephenjaygould.org/ctrl/popper_falsification.html]

      Delete
    10. It does not follow from what Popper says in ref2, but since that's just a short piece, I don't see how that suggests that Popper didn't think it. It really is a pretty big deal for Popper that "corroboration" (his preferred word) is not a form of induction. I.e., well-corroborated theories are not, all else being equal, more likely to be true than poorly-corroborated theories. If I understand you correctly, you think that the Stanford Encyclopedia has just got Popper's views flat out wrong on this point. I'm not sure I share your worries about secondary sources in this context: I'd trust a Popper scholar's interpretation of Popper over my own. I'm on my phone right now so I can't look at Popper's original work, but there is no shortage of other secondary sources saying the same thing. See e.g. page 3-4 of this one: http://www.philsci.com/pdf/BOOKV.pdf

      With regard to falsification meaning something about uniqueness of predictions, I'm not sure I understand what that means.

      Delete
    11. @Alex: I referred to the short piece for the reader's convenience. And because it nicely captures Popper's view.

      Here is the relevant quote from The Logic of Scientific Discovery (Popper 1934):

      "Yet in my view, the whole problem of the probability of hypotheses is misconceived. Instead of discussing the ‘probability’ of a hypothesis we should try to assess what tests, what trials, it has withstood; that is, we should try to assess how far it has been able to prove its fitness to survive by standing up to tests. In brief, we should try to assess how far it has been ‘corroborated’" (pg. 248, Chapter 10, The Logic of Scientific Discovery).

      I hope you will find the above quote more persuasive. I maintain that a lot of secondary sources on Popper are simply mistaken, especially because they try to couch his thinking in a Bayesian framework.

      Even in the lines above, you can see that he has no problem with confirmation ("prove its fitness" or "corroboration"). However, the conditions under which confirmation/corroboration is acceptable are what are crucial for Popper.

      Delete
    12. If anyone wants the context for the quote so that you don't think it was a case of quote mining it, here is the whole paragraph:

      "The attempt has often been made to describe theories as being neither true nor false, but instead more or less probable. Inductive logic, more especially, has been developed as a logic which may ascribe not only the two values ‘true’ and ‘false’ to statements, but also degrees of probability; a type of logic which will here be called ‘probability logic’. According to those who believe in probability logic, induction should determine the degree of probability of a statement. And a principle of induction should either make it sure that the induced statement is ‘probably valid’ or else it should make it probable, in its turn—for the principle of induction might itself be only ‘probably valid’. Yet in my view, the whole problem of the probability of hypotheses is misconceived. Instead of discussing the ‘probability’ of a hypothesis we should try to assess what tests, what trials, it has withstood; that is, we should try to assess how far it has been able to prove its fitness to survive by standing up to tests. In brief, we should try to assess how far it has been
      ‘corroborated’."(pg. 248, Chapter 10, The Logic of Scientific Discovery).

      Delete
    13. @Karthik. We must be talking at cross purposes here. These quotes seem to me to support what I was saying. I.e., that Popper did not think that theories can be confirmed in the usual sense of being shown to be probably true.

      Delete
    14. Just to add, it's a fair point that Popper may not have said that our best scientific theories are probably false if "probably false" means that they are assigned a low probability value according to some inductive principle. But in a more informal sense of "probably", it's in the nature of highly falsifiable theories to be probably wrong.

      Delete
    15. @Alex: You are still looking at it from a Bayesian viewpoint. Outside the Bayesian world, it is not at all clear confirmation is about theories "being shown to be probably true". This is only one view of confirmation. I grant you this is the "common sense" view on the issue too. As I acknowledged (at least, tacitly) earlier, it is true that Popper wasn't a confirmationist in the Bayesian sense. But, Bayesians aren't the authority on confirmation - theirs is just one view of what confirmation means. And to them it is reflected in the posterior probabilities.

      A view before the modern Bayesian view was the view that Pierce discussed in his works. He had a view of confirmation similar to Popper and many others since. i.e. confirmation cannot be about probabilities of hypotheses at all.

      There is in fact a very healthy debate in the literature about what confirmation is. Those who oppose priors on scientific hypotheses (the standard non-bayesian view), hold confirmation is about likelihoods, not posteriors. Popper's spin is that it is about likelihoods provided the theory is falsifiable (i.e., if the facts accord with the theory when the theory makes a special/unique/"falsifiable" prediction, then the theory has received some confirmation.). That Popper meant something different from the normal sense of falsifiability is reasonably clear from the use/discussion of his example of Einstein's theories in the context of Eddington's experiments (this appears to be one of his favorite cases to make his point). As Popper and many after (including David Cox, Mayo) have pointed out, the idea of priors over scientific hypotheses doesn't make sense (except in a few special cases); therefore, posteriors as a reflection of confirmation don't make sense either.

      But, to say Popper is not a confirmationist is to accept the Bayesian view of confirmation as beyond reproach. Especially, when he himself used confirmation as something theories should seek. In this sense, to say Popper is not a confirmationist and to call him a falsificationist is to just miss the nuances of his viewpoint, and to shut down an interesting debate about what such things mean.

      "But in a more informal sense of "probably", it's in the nature of highly falsifiable theories to be probably wrong." Popper disavowed quite strongly any reference to probabilities of hypotheses, as I pointed out. If you have a quote explicitly mentioning Popper discuss how he agrees with informal notions of probably, then it would be nice to see it. But, as I pointed out, the secondary sources that you quoted are simply wrong on the issue of probabilities of hypotheses and Popper's position.

      If this is not clear, I highly recommend reading Deborah Mayo's "Error and the Growth of Experimental Knowledge". She does a wonderful job of clarifying/correcting many misconceptions. I was made aware of my own misconceptions of the issues and enticed into reading the original Popper sources (as opposed to the secondary sources) only because Mayo's writings had initially pointed out that the common view ascribed to Popper as just a falsificationist was plain wrong.

      Also, I think I have started repeating myself, so I shall stop, unless there is a clarification you seek.

      Delete
    16. The relation between falsifiability and probability comes straight from Popper. He did, as you say, reject the notion of the posterior probability of a hypothesis, but he was happy to talk about “logical” (i.e. prior, p. 270) probabilities. Here is what he says on p. 115 of Conjectures and Refutations:

      I explained in my book [=LSD] why we are interested in theories with a high degree of corroboration. And I explained why it is a mistake to conclude from this that we are interested in highly probable theories. I pointed out that the probability of a statement (or set of statements) is always the greater the less the statement says: it is inverse to the content or the deductive power of the statement, and thus to its explanatory power. Accordingly every interesting and powerful statement must have a low probability; and vice versa: a statement with a high probability will be scientifically uninteresting, because it says little and has no explanatory power

      He refers in footnotes to section 83 of LSD, where he says:

      If we compare these views of mine with what is implicit in (inductive) probability logic, we get a truly remarkable result. According to my view, the corroborability of a theory—and also the degree of corroboration of a theory which has in fact passed severe tests, stand both, as it were, in inverse ratio to its logical probability; for they both increase with its degree of testability and simplicity. But the view implied by probability logic is the precise opposite of this. Its upholders let the probability of a hypothesis increase in direct proportion to its logical probability— although there is no doubt that they intend their ‘probability of a hypothesis’ to stand for much the same thing that I try to indicate by ‘degree of corroboration’.

      It’s a bit hard to know what to make of all this. On Popper’s view an ideal scientific hypothesis will be both inherently extremely implausible and also extremely well corroborated. But “corroboration” in Popper’s sense has no probabilistic interpretation. What should someone who is not a total skeptic about epistemic probability conclude from this? The hypothesis is inherently implausible and its high level of corroboration does nothing to counteract that. So, it would seem that the hypothesis is very unlikely to be true. What other conclusion can one reasonably draw? If Popper avoids this conclusion he does so only via a radical and rather implausible skepticism regarding epistemic probability. I think that is what the commentators I referenced were picking up on.

      Delete
    17. This comment has been removed by the author.

      Delete
    18. @Alex: I think I see there is general agreement (better understanding?) here now in most of what Popper meant/said.

      However, the last step of your inferential process is the one that Popper and many others would reject:
      "So, it would seem that the hypothesis is very unlikely to be true."

      This simply doesn't follow without a belief in the utility of posteriors. For (neo-)Popperians, If I understand them correctly, there is just a rejection vs. not vs. corroboration/confirmation (I am being a bit sloppy with this trichotomy. And, it is surely a rather crude partitioning, but cogent given their aversion to posteriors and trustworthy priors).

      As you rightly point out, this implies a "radical" rejection of epistemic probabilities - and to the extent that I can, ascertain that is the correct interpretation of their position. However, it is uncharitable to say it is "rather implausible skepticism" for there is a cogent justification as to why they so radically reject it. Those against the use of epistemic probabilities surely acknowledge such probabilities exist - I think everyone acknowledges they exist at some level. The problem for them is that they don't think one can assess them accurately enough to use them meaningfully in scientific inference. Which is why such talk of posterior probabilities is illegitimate to them.

      Of course it leaves uncertain how one needs to proceed, given that both Bayesian and non-Bayesian philosophies have some justification. But, hopefully, that is a question/problem that philosophers will address and provide more clarity.

      It is good we have come to discussing the validity of epistemic probabilities. Deborah Mayo, who maintains a blog, just blogged about her criticism of epistemic probabilities (http://errorstatistics.com/2014/02/01/comedy-hour-at-the-bayesian-epistemology-retreat-highly-probable-vs-highly-probed-vs-b-boosts/)

      Delete
  2. This comment has been removed by the author.

    ReplyDelete