The Explanatory Filter and Cause of Death

Yes I do deny the utility of using EF to assign cause of death, and I deny coroners reason in a way that parallels the EF.

Nor is this a correct description of their reasoning:

No, the coroner does not rule out the first two before he can be certain of the third.

There is an important fallacy embedded in this type of reasoning. It privileges the last option in cases where there is insufficient information or unreasonable standards of “rule out” are used. This order dependence is well known in machine learning, and a common way that “decision tables” are misunderstood. The EF, in fact, is a type of decision table.

Look at the ordering problem closely.

  1. Is it death by natural cause? If so, conclude accordingly, otherwise move to #2.
  2. Is it death by chance? If so, conclude accordingly, otherwise move to #3.
  3. Conclude death by design.

This is definitely not how coroners determine cause of death. One obvious problem with this approach is that we could just as well have formulated it as:

  1. Is it death by design? If so, conclude accordingly, otherwise move to #2.
  2. Is it death by chance? If so, conclude accordingly, otherwise move to #3.
  3. Conclude death by natural causes.

See the problem here? All of a sudden, we are now biasing inferences towards natural causes, especially in cases where we don’t know from the evidence yet.

So the order of the tests matters to the inference, but there isn’t a way to determine the order of the tests. That is a major short coming.

Likewise, reordering ends up begging the question in a very circular way. For example, how do we determine of something was by natural causes in Dembski’s filter (the first step there)? Well, we could propose using the reordered filter (with natural causes last). But that leaves us with a problem of how to determine of something is designed. So we could propose using Dembski’s filter for this. But then how do we determine whether it was natural causes? And so we fall into an infinite regress.

There are other logical problems with this decision table. It is notable that “unknown” is not an option in this decision table, which demonstrates this isn’t their reasoning, because coroners do determine cause of death to be “unknown”.

So I am an expert in machine learning. One thing we learn very early on is that decision tables like the EF are nothing like how humans reason. They are fundamentally misleading for this reason, because they way we perceive them does not match how they actually work and how they work doesn’t match how we actually reason about the world.

How do we actually determine if it was death by design? We do this (technically it isn’t coroners) by looking for patterns that match what we expect of design, such as evidence of poisons or strangulation marks. It is by weighing much more specified hypothesis of each option against patterns that we observe in the data that we make inferences. In the case of coroners and physicians, there is actually very rigorous guidelines that can be applied to determine what evidence needs to be collected and reported.

Notably, the way we actually reason about design is not order dependent. It does not work like the EF.


Thanks for this clear answer, Joshua.

I like your point about the missing “unknown” option.

Also I see your point about the reordering.

Would you say that in many cases the coroner’s verdict is for the most likely answer, but that the answer is not, strictly speaking, certain?

For example, if the mark on the head is only 95% certain to be caused by a blow with a certain instrument, can the coroner responsibly draw that conclusion? Or if there is even a tiny doubt, must he put down “Unknown”?

1 Like

Dembski’s EF basically operated by asking three questions
(1) Does a law explain it?
(2) Does chance explain it?
(3) Does design explain it?

The EF was found to be fatally flawed years ago because it had no provision for processes which work by a combination of (1) and (2), a combination of laws and chance. Since that is exactly how evolution works Dembski’s EF would always produce false positives of “design” when investigating biological life.

Dembski realized his EF was hopelessly broken and actually announced he would no longer offer the position. Then so many of his rank and file ID-Creationists screamed bloody murder Dembski retracted his retraction. :laughing: Needless to say no one except the most die-hard ID-Creationists even mention the EF anymore.


Coroners should not actually be making final verdicts about design. They are supposed to report what they observe, and state the most likely sequence of events that caused death.

For example, they might say that a body was presented with a rope around the neck, the hyoid bone was fractured, and there were markings on the neck consistent with strangulation, indicating cause of death was asphyxiation. Deeper in the report, much more information would be reported. Alternate causes of death might be discussed.

But there are a lot questions that are very difficult to answer. Who put the rope around the man’s neck? Did the person who did it intend to kill him? Did he intend to kill himself? Was the death accidental or not?

They would put down their best guess, ideally stating certainty in the report, and there are provisions for revising the cause as more information comes in. Note, however, that the coroner is not able to determine if the blow to the head was intentional or not, or if it was intended to kill. That inference is far more difficult, and it isn’t possible to do by merely looking at a body.

Instead coroners are identifying the proximate causes of death. They do determine if it is most likely homicide or suicide, but that is not a final determination.

The bigger question isn’t how coroners determine the intention to cause death, but how courts determine it and how we would infer it as people observing the facts. It should become clear that the details matter. Different theories might be considered and rejected based on their consistency with the data. Often, we may not be able to determine conscious intent, because that isn’t directly visible.

It is notable that the legal system has ways of determining guilt that do not require conscious intent. That is required because private intent often (usually?) cannot be determined with much certainty. Instead, we can sometimes see the consequence of particular actions (whether intended or not), and judge behavior based on what a reasonable person would do in the same situation. We can’t really judge the private intentions.

[NOTE: I edited this to be more accurate about coroner’s determinations after some research. My initial response was not accurate.]


I don’t think that was the problem. Dembski had no way of calculating what would be the probability of finding an adaptation that good (or better) under ordinary evolutionary processes. He called for this calculation but left it to the user to figure out how to do it. Then he uses that probability, which the user has somehow come up with, to calculate whether there is CSI. If the probability is low enough, we have CSI. But basically what we are trying to compute is whether the probability is implausibly low. So we, the users, have to do all the heavy lifting, and the declaration of CSI being present is just after-the-fact window-dressing.

1 Like

Isn’t that at least one problem with the EF?

But that also makes it convenient that he doesn’t actually need this calculation for the EF to work if we follow it as written.

I suppose that is one interpretation, but doesn’t he claim to actually apply the EF?

How does he do the step where he asks whether “necessity” can account for the specification? I think he needs the probability calculation there, or asks you to supply it?

He has some calculation for the flagellum, which I think is basically from Behe, and does not allow for natural selection to put together any precursor. But in all other cases involving evolution, I don’t know how he does any probability calculation.

It seems to me (after years of reading Dembski) that one could make a version of the EF that is totally consistent with using CSI. One needs to (1) to alter his EF to not separate “chance” and “necessity” but to replace them by “ordinary evolutionary and chemical and physical processes”, and (2) make sure that people understand the can-we-explain as not assessing complete logical impossibility but instead just implausibly low probability, an event so improbable that it is expected to occur less than once in the history of the observable universe. Then his 2005 SC > 0 calculation is basically the same as that (modified) Explanatory Filter.

1 Like

Need we mention: A a coroner might infer that death occurred due to actions of another human or other natural cause, but would not infer the existence of something otherwise unknown.


Yes that seems about right @Joe_Felsenstein . There are several key points though.

  1. This is a modified EF, not the one he proposed.

  2. It would be known, not merely ordinary, evolutionary and chemical and physical processes.

  3. The modified filter still would assign the products of unknown and difficult to model evolutionary processes to design, rather than concluding “unknown” (as would be justified).

  4. CSI is still subject to the sharpshooter fallacy, for a set of reasons covered many many times.

  5. This modified EF still has the ordering problem discussed above. The order of execution matters, there isn’t a principled way to determine order, and it begs an infinite regress.

So, this modified EF still fails, showing that even attempts to modify Demski’s EF don’t rescue it.

It strikes me also that this modified filter seems to echo a frequentist approach to hypothesis testing in a very particular way. Rather than assessing if correlation is significant, it is assessing whether or not there is unexplained variance, even if that variance is a very low percentage of total variance. That unexplained variance is determined to be design, but could just as well have been random noise that we cannot account for in modeling.

In this case, a Bayesian approach would be better. I don’t think Bayesian reasoning is a panacea, but it avoids all these particular pitfalls. Rather than determining if there is unexplained variance (there is), the Bayesian approach would weight different well-specified explanations against each other, requiring IDists to produce a well specified explanations (design principles?). A Bayesian approach is also mathematically commutative, so it does not suffer from the ordering problem.

1 Like

This unexplained variance also occurs when you apply an incorrect model, and Demski never considers anything resembling an evolutionary model.

1 Like

@Josh, note that I was not advocating for either the EF, the modified EF, or CSI / SC. I was just arguing that the modified EF could be made equivalent to using SC. As for the Texas Sharpshooter issue, it is less of a problem if fitness is the criterion for the scale on which specification is measured.

It occurs to me that the original EF does not talk about such a scale, does not talk about the probability of getting a result as good as observed, or better. CSI / SC does use such a scale. So maybe that is one big difference. I think that the scale is actually there, implicitly, in the EF but it this seems not to be discussed by Dembski.

1 Like

In one of his formulations Dembski defines a “critical region” in the sense of a statistical hypothesis test. It must be from one of Dembski’s books, but I’ve never figured out which one.

1 Like

In one of his CSI / SC explanations? Or in an account of the EF?

1 Like

I’ve seen it Copy/Pasta in terms of CSI/SC, or at least I do not recall any mention of EF in the excerpt. I never found it in Dembski’s published papers, therefore I assume it’s from a book.

OK, CSI / SC is defined in terms of a scale of specification. It’s EF where I don’t recall a scale being explicitly invoked. (Nevertheless I think it is implicit).

I’ll need to read up on the EF again, it never made much sense to me in the first place. An implicit scale might make sense given uncertainty in the decisions at each step.

I recently had the idea of looking into type I and II error rates for CSI/SC. There is no such thing as a perfect test, and this is not address by Dembski or anyone in ID. Outside of ID I doubt anyone thinks this is worth the trouble, but I kind of like tilting at windmills. :slight_smile:

The problem with investigating Type I and Type II errors is that you would have to know the null distribution under natural evolutionary processes. And that is what we have had so much trouble discerning in Dembski’s work.

You might look up the recent paper by George Montañez in the Biologic Institute’s journal BIO-Complexity. There he tries to unify all the forms of Specified Information and Functional Information, using a uniform distribution over all genetic sequences as the null distribution, and taking the specification function and renormalizing it by dividing by its sum to get an alternative probability distribution. He then applies standard hypothesis testing machinery. But he never discusses why using the uniform distribution is appropriate as the outcome of evolutionary processes.


I would work it out with Dembski’s uniform assumption (and independence), then look at power with respect to non-uniform distribution and non-independence. I’ll let you know if I come up with anything interesting.

I will, thanks!

Here is something - Any distribution other than IID uniform will have less Shannon Information and will be more predictable (will not generate the full range of independent sequences). This might require adjusting the Phi function to account for less information in the population.

Dembski’s is very sensitive to correlation (non-independence), and if not accounted for will increase CSI. The correlation between bits must be effectively zero, or rapidly decay to zero as the distance between bits in the sequence increases. If the distribution is known and correlation weak and decays rapidly, then this can be corrected. In DNA data Linkage Disequilibrium is correlation which decays with the distance between loci, and will increase CSI if not corrected.

For those not familiar with the math; the uniform distribution that Dembski assumes carries the maximum information for any sequence of a given length. Non-uniform distributions and predictable sequences (correlation) carry less information, and may give the false appearance of CSI.