Saturday, November 15, 2008

Personal Bias and Peer Disagreement

The epistemic issues that arise when considering how to rationally respond to evidence of one's own bias are, I think, exactly the same as those discussed in the peer disagreement literature. The core issue, in both cases, is how to weigh higher-order evidence (of our own possible unreliability) against the first-order evidence. If I believe P when an epistemic peer judges my evidence to instead support not-P, this seems isomorphic to a case where I know 50% of people in my position will be mistaken due to bias [corrected, thanks Daniel]. So the same three basic positions suggest themselves in either case:

(1) Higher-order evidence trumps. This is called the 'Equal Weight View' in the peer-disagreement case. In the personal bias case, we might call it the 'Automatic Adjustment View', since it claims you should adjust your initial judgment to compensate for merely possible bias, no matter what the first-order evidence actually recommends. This is essentially the claim that you're in the same epistemic position whether you turn out to be actually biased or not. (After all, the argument goes, from the inside you can't tell which of the two positions you're in. So any rule you can follow will have to treat both situations the same.)

(2) First-order evidence trumps. In the peer disagreement case, this is called the 'Asymmetric No Independent Weight View' -- since the thought is that whoever is actually right about what the first-order evidence supports may thereby "stick to their guns", giving no weight to the opinion of their mistaken peer, but the mistaken person should of course change their opinion to what the evidence really supports. (And too bad if they don't realize this.) This is essentially the claim that learning of possible bias doesn't change your epistemic situation; either the evidence supports P or it doesn't, and your own (in)capacities don't change this fact.

(3) The Total Evidence View. Finally, we might think one is rationally required to take all the evidence into account. Neither first- nor higher-order evidence necessarily trumps the other; though in particular cases it may do so. Note that the inclusion of the first-order evidence introduces some 'asymmetry', but even the correct person may be required to compromise their initial judgment somewhat in light of the (misleading, as it happens) higher-order evidence of their unreliability.

Note that it would seem absolutely bizarre for someone to hold different views about peer disagreement and personal bias. If you think that first-order evidence trumps in case of peer disagreement, it would be terribly inconsistent to suddenly turn around and say the higher-order evidence trumps in case of personal bias. The two issues go hand-in-hand. I'll wrap up this post by noting two further points of correspondence:

(A) The 'equal weight' and 'automatic adjustment' views both gain their plausibility from considering perceptual (more precisely: non-inferential) examples. But this is misleading, because we're effectively imagining a case where there is no (other) first-order evidence; the judgment itself is all we have to go on, so of course unopposed higher-order evidence about the reliability of this judgment is going to be decisive.

(B) Any view which disregards the epistemic importance of first-order evidence is decisively refuted by the problem of implausibly easy bootstrapping. Here I adapt Tom Kelly's objection to the EWV to instead apply to Erica Roedder's automatic adjustment view:
Suppose I receive an incoherent C- paper, and irrationally misjudge it to deserve an A. If the author happens to be a black student and I have evidence that instructors are typically biased against black students, does that automatically mean that all things considered I should boost the grade and give it an A+? Surely not. My initial judgment may be some evidence that the paper warrants an A+. But it's not the only relevant fact. In particular, it can't make up for the fact that the details of the paper itself constitute overwhelming evidence that the paper should be given a C-. The presence of higher-order evidence doesn't excuse me from concluding what the first-order evidence demands.

In short, the AAV (if understood as making positive claims about what I rationally ought to judge) makes it too easy for me to get things right. No matter what loony judgment I initially make, if I then amend it slightly in response to higher-order evidence of bias, then my resulting judgment is guaranteed to be rationally justified! Too easy.

P.S. For background, see Tom Kelly's excellent paper: 'Peer Disagreement and Higher Order Evidence' [pdf].


  1. I think this is right, although I'm not sure if I do for the same reasons; I think both the personal bias and peer disagreement cases are instances of issues that arise when dealing with checking mechanisms, i.e., the means by which we check and double-check that our conclusions are right: namely, those issues that have to do with cases where a (non-ideal) checking mechanism registers us as (probably) wrong despite our not having in hand anything (other than the information of the checking mechanism itself) that suggests that we really were.

  2. It seems to me that there's a way of stating the equal weight view (or something like it) that doesn't make it equivalent to the "higher order evidence trumps" view.

    You might state a version of the equal weight view along the following lines: "insofar as you ought to think that your peer is just as reliable as you are at judging the truth value of propositions like P, then you ought give your peer's view about whether P the same weight you give your own."

    It's consistent with this view that sometimes, depending on the circumstances, learning about a disagreement with someone makes it reasonable to no longer regard that person as your peer. (consider the restaurant cases David Christensen discusses, where your dinner companions, attempting to calculate an 18% tip, come up with an answer many times larger than the total bill. Perhaps it was reasonable to regard them as your peers with respect to arithmetic prior to hearing their answer, but not after).

    We might put this by saying that in cases like these, first order evidence introduces a relevant asymmetry such that you oughtn't think that the person who you previously regarded as your epistemic peer is, at least with respect to this question, your peer.

    But if the first order evidence doesn't do that--if it leaves it most reasonable for you to think that she's just as reliable as you are--then you ought to give her view the same weight that you give your own.

    I don't think this trivializes the view--it still has the perhaps counterintuitive implication that it's never reasonable to regard somebody else as just as reliable as you with respect to questions like P, while also thinking that you're more likely to right about P. It's not clear that Kelly's view has this implication, and it's certainly clear that views on which it's permissible to give your own view extra weight just because it's yours don't.

  3. One more thing

    It would be very strange to have a view on which it could be reasonable to both:

    1. have a credence of 0.5 that my bias led me to judge incorrectly when I initially judged that p


    2, have a credence of > 0.5 that p.

    If you satisfy 1 and 2 you are probabilistically incoherent. If you satisfy 1, you must have a credence of at least 0.5 that p is false, since the proposition that you judged that p incorrectly implies that p is false. And it's clearly impossible to have a credence of at least 0.5 that p is false, and also satisfy 2 while being probabilistically coherent.

    It seems to me that insofar as you think your first order evidence changes what credence it's reasonable to have in p, it should also change what credence its reasonable to have in the proposition that your bias led you to falsely judge that p. I'm not saying you denied that, but at least one (probably uncharitable) way of reading the total evidence view is that it implies that your first order evidence affects what it's reasonable to believe about whether p, without affecting what it's reasonable to believe about whether your bias led you to falsely judge that p.

  4. Thanks Daniel, I was a bit sloppy there. Rather than "50% credence", I should have said "50% of people in situations like mine are led by bias to a false conclusion".

    To clarify your first comment, are you talking about revising our estimation of the other person's general reliability (on questions relevantly like the one under consideration), or just their accuracy in this specific case? If the latter, the view seems (relatively) trivial. The former view strikes me (prima facie) as an unstable compromise, but I'd need to think about it more.

    Incidentally, I think Adam Elga's take on the Christensen case is that it's not really the first-order evidence we (assuming we're Equal Weight theorists) take into account. Rather, it's the higher-order asymmetry that (only) one party finds the other's calculation obviously absurd.

  5. In response to your clarification question, I do think it's tricky to walk between the Scylla of triviality and the Charybdis of falsity here.

    I'd like to say that it's the other person's general reliability on questions relevantly like the one under consideration, but it's not clear how to understand "relevantly like." Obviously, in cases where you know that the other person is very drunk, there's some sense in which you can think they're generally just as reliable as you are about questions like P, but in which you can also reasonably think that you're more likely to be right in this particular disagreement. The equal weight view had better not say there's something wrong with that.

    The hope would be that in each case, you can find some level of generality at which you understand the phrase "relevantly like" such that if I think I'm more likely to be right than you about P, I must think I'm more likely to be right in general about questions relevantly like whether or not P. But this level of generality must not be the trivial one where the only question relevantly like whether or not P is the question of whether or not P.

    I don't think this is so much an objection to the view, as it is a reason to think there isn't even a good way to formulate it. Though this may also a problem for the total evidence view. If we characterize higher order evidence with respect to p as something like evidence about how likely I am to be right about questions relevantly like whether p, it looks like Kelly faces similar issues. Maybe formulating the debate requires solving something like the generality problem. That'd be depressing.

  6. Do we need to talk about generality? Here's Elga's view (from 'Reflection and Disagreement', p.17): "Upon finding out that an advisor disagrees, your probability that you are right should equal your prior conditional probability that you would be right. Prior to what? Prior to your thinking through the disputed issue, and finding out what the advisor thinks of it. Conditional on what? On whatever you have learned about the circumstances of the disagreement."

    Basically, we abstract away the disputed first-order evidence, and whatever is left (including facts about drunkenness, confident feelings that the other has made an absurd mistake, etc.) counts as higher-order evidence. It would seem a strange kind of 'equal weight view' that went beyond these higher-order features, allowing first-order details to override the prior conditional probability Elga describes here.

  7. The suggestion that we should "abstract away the disputed first order evidence", and that this will help us avoid the generality issues presupposes that we can make the distinction between first order and higher order evidence without dealing with the generality issues.

    Here's how the worry applies to the Elga view. When we find that we disagree, we learn a lot. He needs some way of saying which parts of what we learn go into the "circumstances of disagreement", and which don't. And the stuff that goes into the circumstances of disagreement must be less than our total new evidence, because if it weren't his view would collapse into the plausible but not novel dictate that we should conditionalize on what we learn. So the way the generality issues seem to arise for Elga is that he has to decide how much to abstract away from the particular case to get to the "circumstances of disagreement."

    I agree that in many particular cases, it's easy to get a handle on what looks like it should count as first order evidence that we can abstract away from, and what doesn't (and in particular, that facts about drunkenness, confident feelings that the other has made an absurd mistake, etc. seem like they shouldn't count as first order evidence). But it's not clear to me that we have a general understanding of the distinction. That's what my earlier post was getting at--if the distinction is just that the first order evidence is everything relevant to whether p, and the higher order evidence is everything relevant to how reliable one is at answering questions relevantly like p, then we've got the generality worries. (Also, first order evidence and higher order evidence certainly don't look mutually exclusive if we characterize things that way...)

    Maybe we don't need a general characterization of the distinction between first order and higher order evidence--maybe we have a good enough handle on how it applies in particular cases to formulate and compare views that use the distinction. But you can't avoid the generality worries by suggesting that we understand what the higher order evidence is by abstracting away from first order evidence and taking what's left--if you have a general account of what that involves, you've already solved the problem.

  8. Alas for lack of time coordination - at the moment I'm too busy with other issues to discuss this properly, and when I'm in the mood for this I suppose your mind will be elsewhere. :)

  9. Richard,

    Two things:
    1. Why think in cases of perception there is no first order evidence? This seems really strange. Why not think that the perceptual experiences are first-order evidence. These do no equal the judgment.
    2. I don't follow your adapted Kelly objection. No evidentialist should want to claim that one piece of evidence evidence can affect what one should 'all things considered' believe. But, this is not to say that this is in virtue of first=0rder evidence. I would insist that it is due to other higher-order evidence you have. More needs to be said to convince me that it is the first-order evidence that is doing any work.

  10. Robin - I'm happy to return to this topic if you comment again sometime.

    Jon - (2) what "other higher-order evidence" is there in the case I describe? I meant to describe a case in which there is none -- at least none that tends in the right direction, towards my giving a lower rather than higher grade -- so the epistemic requirement must be due to the first-order evidence.

    (1) By 'perceptual judgment' I mean the judgment that is implicit in the perceptual experience itself. So, sure, the perception itself is first-order evidence, but it is undermined by my learning that I have a perceptual bias: a line's appearing to me as though 1 inch long is really evidence that it is 2 inches long. The relevant point, for my purposes, is that there is no other first order evidence besides the undermined perception.

  11. I was taking it that in the case you had reasons to think that you hadn't been so biased regarding this paper. I think that if you got evidence that you had been biased regarding this paper, and didn't have any defeater for that, then you should bump up the grade. This would be so, no matter how bad the paper was (no matter how strong the first order evidence was).

  12. That's the bootstrapping case: by stipulation, there is no defeater for the higher-order evidence (i.e. no special reason to think that I'm one of the few instructors to lack some degree of implicit racial bias). So you think that in the described case, all things considered, I should give the incoherent paper an A+.

    And more generally: No matter what loony judgment I initially make, if I then amend it slightly in response to higher-order evidence of bias, then my resulting judgment is guaranteed to be rationally justified!

    I do not consider that a remotely plausible position, but your mileage may vary.

  13. where does Kelly give this case?

  14. It's my case. Kelly discusses bootstrapping in relation to the Equal Weight View starting p.15 in the linked pdf.


Visitors: check my comments policy first.
Non-Blogger users: If the comment form isn't working for you, email me your comment and I can post it on your behalf. (If your comment is too long, first try breaking it into two parts.)

Note: only a member of this blog may post a comment.