Winston Ewert has a post at Evolution News & Views that directly responds to my post here, A CSI Challenge which is nice. Dialogue is good. Dialogue in a forum where we can both post would be even better. He is extremely welcome to join us here
In my Challenge, I presented a grey-scale photograph of an unknown item, and invited people to calculate its CSI. My intent, contrary to Ewert’s assumption, was not:
…to force an admission that such a calculation is impossible or to produce a false positive, detecting design where none was present.
but to reveal the problems inherent in such a calculation, and, in particular, the problem of computing the probability distribution of the data under the null hypothesis: The eleP(T|H)ant in the room
In his 2005 paper, Specification: the Pattern that Signifies Intelligence, Dembski makes a bold claim: that there is an identifiable property of certain patterns that “Signifies Intelligence”. Dembski spends the major part of his paper on making three points:
- He takes us on a rather painful walk-through Fisherian null hypothesis testing, which generates the probability (the “p value”) that we would observe our data were the null to be true, and allows us to “reject the null” if our data fall in the tails of the probability distribution where the p value falls below our “alpha” criterion: the “rejection region”.
- He argues that if we set the “alpha” criterion at which we reject a Fisherian null as 1/[the number of possible events in the history of the universe], no way jose will we ever see the observed pattern under that null. tbh I’d be perfectly happy to reject a non-Design null at a much more lenient alpha that that.
- He defines a pattern as being Specified if it is both
- One of a very large number of patterns that could be made from the same elements (Shannon Complexity)
- One of a very small subset of those patterns that can be defined as, or more simply than the pattern in question (Kolmogorov compressibility)
He then argues that if a pattern is one of a very small specifiable subset of patterns that could be produced under some non-Design null hypothesis, and that subset is less than 1/[the number of possible events in the history of the universe] of the whole set, it has CSI and we must conclude Design.
The problem, however, as I pointed out in a previous post, Belling the Cat, is that the problem with CSI is not computing the Specification (well, it’s a bit of a problem, but not insuperable) nor with deciding on an alpha criterion (and, as I said, I’d be perfectly happy with something much more lenient – after all, we frequently accept an alpha of .05 in my field (making appropriate corrections for multiple comparisons) and even physicists only require 5 sigma. The problem is computing the probability of observing your data under the null hypothesis of non-Design.
Ewert points out to me that Dembski has always said that the first step in the three-step process of design detection is:
Identify the relevant chance hypotheses.
Reject all the chance hypotheses.
and indeed he has. Back on the old EF days, the first steps were to rule out “Necessity” which can often produce patterns that are both complex and compressible (indeed, I’d claim my Glacier is one) as well as “Chance”, and to conclude, if these explanations were to rejected, Design. And I fully understand why, for the sake of algebraic elegance, Dembski has decided to roll Chance and Necessity up together in a single null.
But the first task is not merely to identify the “relevant [null] chance hypothesis” but to compute the expected probability distribution of our data under that null, which we need in order to compute the the probability of observing our data under that null, neatly written as P(T|H), and which I have referred to as the eleP(T|H)ant in the room (and, being rather proud of my pun, have repeated it in this post title). P(T|H) is the Probability that we would observe the Target (i.e. a member of the Specified subset of patterns) given the null Hypothesis.
And not only does Dembski not tell us how to compute that probability distribution, describing H in a throwaway line as “the relevant chance hypothesis that takes into account Darwinian and other material mechanisms”, but by characterising it as a “chance” hypothesis, he implicitly suggests that the probability distribution under a null hypothesis that posits “Darwinian and other material mechanisms” is not much harder to compute than that in his toy example, i.e. the probability distribution under the null that a coin will land heads and tails with equal probability, in which the null can be readily computed using the binomial theorem.
Which of course it is not. And what is worse is that using the Fisherian hypothesis testing system that Dembski commends to us, our conclusion, if we reject the null,is, merely that we have, well, rejected the null. If our null is “this coin is fair”, then the conclusion we can draw from rejecting this null is easy: “this coin is not fair”. It doesn’t tell us why it is not fair – whether by Design, Skulduggery, or indeed Chance (perhaps the coin was inadvertently stamped with a head on both sides). We might have derived our hypothesis from a theory (“this coin tosser is a shyster, I bet he has weighted his coin”), in which case rejecting the null (usually written H0), and accepting our “study hypothesis” (H1) allows us to conclude that our theory is supported. But it does not allow us to reject any hypothesis that was not modelled as the null.
Ewert accepts this; indeed he takes me to task for misunderstanding Dembski on the matter:
We have seen that Liddle has confused the concept of specified complexity with the entire design inference. Specified complexity as a quantity gives us reason to reject individual chance hypotheses. It requires careful investigation to identify the relevant chance hypotheses. This has been the consistent approach presented in Dembski’s work, despite attempts to claim otherwise, or criticisms that Dembski has contradicted himself.
Well, no, I haven’t. I’m not as green as I’m cabbage-looking. I have not “confused the concept of Specified Complexity with the entire design inference”. Nor, even am I confused as to whether Dembski is confused. I think he is very much aware of the eleP(T|H)ant in the room, although I’m not so sure that all his followers are similarly unconfused – I’ve seen many attempts to assert that CSI is possessed by some biological phenomenon or other, with calculations to back up the assertion, and yet in those calculations no attempt has been made to computed P(T|H) under any hypothesis other than random draw. In fact, I think CSI, or FCSI, or FCO are a perfectly useful quantities when computed under the null of random draw, as both Durston et al (2007) and Hazen et al 2007 do. They just don’t allow us to reject any null other than random draw. And this is very rarely a “relevant” null.
It doesn’t matter how “consistent” Dembski has been in his assertion that Design detection requires “careful investigation to identify the relevant chance hypothesis”. Unless Dembski can actually compute the probability distribution under the null that some relevant chance hypothesis is true, he has no way to reject it.
However, let’s suppose that he does manage to compute the probability distribution under some fairly comprehensive null that includes “Darwinian and other material mechanisms”. Under Fisherian hypothesis testing, still, all he is entitled to do is to reject that null, not reject all non-Design hypotheses, including those not included in the rejected “relevant null hypothesis”.
Ewert defends Dembski on this:
But what if the actual cause of an event, proceeding from chance or necessity, is not among the identified hypotheses? What if some natural process exists that renders the event much more probable than would be expected? This will lead to a false positive. We will infer design where none was actually present. In the essay “Specification,” Dembski discusses this issue:
Thus, it is always a possibility that [the set of relevant hypotheses] omits some crucial chance hypothesis that might be operating in the world and account for the event E in question.
The method depends on our being confident that we have identified and eliminated all relevant candidate chance hypotheses. Dembski writes at length in defense of this approach.
But how does Dembski defend this approach? He writes
At this point, critics of specified complexity raise two objections. First, they contend that because we can never know all the chance hypotheses responsible for a given outcome, to infer design because specified complexity eliminates a limited set of chance hypotheses constitutes an argument from ignorance.
In eliminating chance and inferring design, specified complexity is not party to an argument from ignorance. Rather, it is underwriting an eliminative induction. Eliminative inductions argue for the truth of a proposition by actively refuting its competitors (and not, as in arguments from ignorance, by noting that the proposition has yet to be refuted). Provided that the proposition along with its competitors form a mutually exclusive and exhaustive class, eliminating all the competitors entails that the proposition is true.
But eliminative inductions can be convincing without knocking down every conceivable alternative,a point John Earman has argued effectively. Earman has shown that eliminative inductions are not just widely employed in the sciences but also indispensable to science.
Even if we can never get down to a single hypothesis, progress occurs if we succeed in eliminating finite or infinite chunks of the possibility space. This presupposes of course that we have some kind of measure, or at least topology, on the space of possibilities.
Earman gives as an example a kind of “hypothesis filter” whereby hypotheses are rejected at each of a series of stages, none of which non-specific “Design” would even pass, as each requires candidate theories to make specific predictions. Not only that, but Earman’s approach is in part a Bayesian one, an approach Dembski specifically rejects for design detection. Just because Fisherian hypothesis testing is essentially eliminative (serial rejection of null hypotheses) does not mean that you can use it for eliminative induction when the competing hypotheses do not form an exhaustive class, and Dembski offers no way of doing so.
In other words, not only does Dembski offer no way of computing the probability distribution under P(T|H) unless H is extremely limited, thereby precluding any Design inference anyway, he also offers no way of computing the topology of the space of non-Design Hypotheses, and thus no way of systematically eliminating them other than one-by-one, never knowing what proportion of viable hypotheses have been eliminated at any stage. In other words, his is, indeed, an argument from ignorance. Earman’s essay simply does not help him.
Suffice it to say, by refusing the eliminative inductions by which specified complexity eliminates chance, one artificially props up chance explanations and (let the irony not be missed) eliminates design explanations whose designing intelligences don’t match up conveniently with a materialistic worldview.
The irony misser here, of course, is Dembski. Nobody qua scientist has “eliminated” a “design explanation”. The problem for Dembski is not that those with a “materialistic worldview” have eliminated Design, but that the only eliminative inductionist approach he cites (Earman’s) would eliminate his Design Hypothesis out of the gate. That’s not because there aren’t perfectly good ways of inferring Design (there are), but because by refusing to make any specific Design-based predictions, Dembski’s hypothesis remains (let the irony not be missed) unfalsifable.
But until he deals with the eleP(T|H)ant, that’s a secondary problem.
Edited for typos and clarity