Proof Probability & Statistics & AI Don't Discover Cause

Everybody has read, or seen adaptations of, Bram Stocker's documentation of Dracula, the undead count, unlate of Transylvania. From these pages and films, we all know how to kill a vampire. In Van Helsing's own words, "I shall cut off her head and fill her mouth with garlic, and I shall drive a stake through her body."

Nipping off a noggin in plain enough, as is stuffing a mouth full of the stinking rose. But there is still ambiguity: what kind of stake?

Again, Van Helsing: "a round wooden stake, some two and a half or three inches thick and about three feet long."

Important information, yet Stoker still failed to tell us what kind of wood. Luckily, Charles Dickens, a mentor and confidant of Stoker, had Scrooge give us the answer, an answer that was plain to contemporaries of both men, but which plainness has faded somewhat for us. Scrooge said, "every idiot who goes about with 'Merry Christmas' on his lips, should be boiled with his own pudding, and buried with a stake of holly through his heart. He should!"

So it's holly. The stouter branches of ilex aquifolium, to be exact, that plant being indigenous to Europe.

But what if we tried maple, oak, or cherry stakes? Pine? Apple? Stoker, and Dickens with him, didn't absolutely say it had to be holly, or that it couldn't be any other wood. Or that the stake couldn't be 3.1 inches thick, or only 2.5 feet long.

Or what color. Or of green wood or dried.

Obviously, we must experiment! Are we not told that the only sure path to scientific knowledge---which is like regular knowledge, but better, because it is certified by Experts---is the randomized controlled experiment? Do they not insist that anecdotal evidence, the data gathered by our own eyes and experience, does not count?

Therefore we must plan a randomized controlled experiment, calculate a p-value, and then, and only then, can we decide if other woods, lengths, colors, smells, thicknesses, pliabilities, bark status, left- or right-handed thrusts, stakes with and without vampire heads lopped off, head lopped off first then stakes through heart, and vice versa, black versus white garlic, raw versed roasted, and so on and so forth.

We need to gather a large number of undead, the exact number being decided by complex sample size equations, thrusting some through with holly, others with oak, and so on, all randomized. If we don't randomize, the results we gathered would not count.

Why? They would not have the blessing on randomization upon them, a blessing which is needed in order to apply probability to the results. Or, rather, to extract it. The results themselves are imbued with probability by the magic of randomization.

If this sounds absurd to you, then you have not had statistical training, and if you did, it didn't take, for which you may consider yourself lucky.

The rules are that randomized controlled trials must be conducted and that other evidence, while it can be considered, with a sniff, the nose arched, will never be conclusive. P-values---those are conclusive.

When I wrote Uncertainty, ISIS was making their tour in the Mideast. They were then crucifying their enemies. If memory serves, their victims were crucified on the standard wooden posts and crosses. Old fashioned, but effective.

I wondered what would happen if we switched from wood to metal. Obviously, using null hypothesis significance theory, or even standard Bayesian theory, we'd have no choice but to conduct "gold standard" randomized controlled experiments. Our prejudice that metal would serve just as well as wood just would not do. Only "data" generated in the appropriate matter can count.

Now, even a strict frequentist would agree that the substance of the crucifixion---wood, metal, or whatever---would not matter. What does is the exposure and rough treatment. These kill, not the cross itself. That conclusion is deduced. Which is a violation of statistical theory.

That's not rare. All frequentists, for instance, always and without fail interpret confidence intervals and p-values as Bayesians would, a strict error according to their theory. This never seems to bother them, even as they defend their theory.

It should be clear we are always after cause, even if we can't reach it. It should also be clear we do not need to conduct new experiments to deduce conclusions about cause. I stress "new" because we already have gained substantial knowledge of steel and iron, for instance, through other means, which allows us to deduce metallic crucifixion would work swell. We only conduct experiments in an attempt to confirm our suspicions of cause.

Subscribe or donate to support this site and its wholly independent host using credit card or PayPal click here