More Thoughts on Falsifiability

Consider the concept of meta-attraction. It is the assertion that many trans women are attracted to men not because of normal androphilia, but instead because being with men makes them feel feminine, which they find attractive because of their autogynephilia. Skeptics of Blanchard’s typology seem to find this theory preposterous and often use the mere fact that this has been proposed to dismiss the typology.

This seems to imply a strong belief that meta-attraction is an incorrect theory, which makes me tempted to suggest an experiment. The experiment would be pretty simple; take some bisexual-identifying autogynephilic trans women, show them pictures of men or women, and measure their genital arousal. If they exhibit a gynephilic arousal pattern, we declare that meta-attraction is the correct theory, gender identity theory has been debunked, and the typology is correct. After all, gender identity theory clearly very strongly predicts that bisexual-identifying trans women are bisexual in an ordinary sense, so this should falsify that model.

Here’s the problem with this, though: if they did end up having a bisexual arousal pattern, my belief in the typology wouldn’t be all that shaken. Sure, I would probably adjust some expectations a bit, but I’d only rework things a little. The reason for this is that I can imagine a number of ways that AGP trans women might still exhibit genital arousal even if they are only meta-attracted. For example, maybe the genital arousal is learned, maybe they can manage to construct fantasies powerful enough to awake their meta-attraction using only the images, or maybe being paraphilic can lower your sexual specificity. Point is, there’s a lot of possibilities.

But wait, doesn’t this violate the principle of falsifiability? Or the law of conservation of expected evidence? Or some other basic rule of epistemology? Well, the great thing is that we have math that tells us how to form accurate beliefs, so we can see what this says. Let’s build a toy model:

Suppose we say that the only two possible theories are Blanchard’s theory, which we denote B, and the gender identity theory, which we denote I. We then consider the observation E that trans women will exhibit gynephilic arousal patterns. Due to the previously mentioned complexities, I am uncertain, so I will perhaps say P(E|B) = 0.7. Gender identity theory, on the other hand, confidently asserts that bisexual trans women are not meta-attracted but instead bisexual in a traditional sense, so we might say that P(E|I) = 0.05. Let’s assume an equal prior probability of P(B) = P(I) = 0.5. In that case, we can find the prior probability of E, namely P(E|B)P(B)+P(E|I)P(I) = 0.375. Discovering that trans women exhibit gynephilic arousal patterns yields a posterior of P(B|E) = P(E|B)P(B)/P(E) ~ 0.93, and therefore P(I|E) ~ 0.07, clearly moving the probabilities a lot. On the other hand, discovering that trans women exhibit bisexual arousal patterns yields a posterior of P(B|not E) = P(not E|B)P(B)/P(not E) = 0.24, which is more than three times as high probability as the gender identity model would be assigned if we had found gynephilic arousal patterns.

(Of course, this toy model is very incomplete. For example, I don’t think either side would accept the idea that the two theories are currently equally supported by evidence, and the toy model pretends that both sides are honest truthseekers, a proposition that can easily be shown to be wrong by applying something similar to Aumann’s agreement theorem.)

So, first, doesn’t this violate the scientific rule that a scientific finding can only be considered scientific evidence for a scientific proposition if the scientific experiment would have disproved the proposition if it had turned out differently? Well, first of all, I’m not even sure this is a rule of science (but people keep accusing me of violating it and throwing around the word “science”, so it’s at least a rule of the popular conception of science), and even if it is, that doesn’t necessarily mean that it’s a valid rule for figuring out the truth (that rule would be Bayes’s theorem). In fact, this rule clearly doesn’t hold. It both fails to hold in a very strict sense (the rule is formulated in absolute rather than quantitative language, which is an easy way to see that it is invalid because beliefs are not binary) and in a more practical sense. In the simplest example, if someone tells you that some event has happened, then it probably has happened, but that doesn’t mean it hasn’t happened if nobody has told you; it might just be unimportant enough that you haven’t heard about it.

(But what about the law of conservation of expected evidence? Well, first, you’ll notice that the evidence did in fact update away from the typology. However, since the expectation overall was that we wouldn’t even find the evidence in the first place (P(E) = 0.375 < 0.5), we didn’t need to update as far away from the theory when finding it as we update towards the theory when not finding it. Of course, P(E) depends on the prior beliefs about the theory, so when P(B) > 0.69, the asymmetry switches the other way around.)

One thing that’s worth noting here is that it ties in with confirmation bias. “Confirmation bias” is usually defined as only seeking evidence in favor of one’s theory and dismissing evidence against. I think there’s an important distinction here, in that my approach is (or at least, can be) perfectly rational, whereas confirmation bias by definition is irrational. More specifically, if I already know the outcome of an experiment, it’s a waste of time to perform it; technically it would be good as a sanity check, but I could instead be focusing on something where I don’t know the results. This means that it’s a waste of time to perform experiments that my current model can very confidently predict the results of. At the same time, for social reasons it’s useful to perform experiments where other people are very confident about the results and I’m not, as this will give me a ton of relevant information and arguments. This means that implicitly, it’s rational for me to do experiments where my model doesn’t get disproved by the results no matter what the outcome is, but other’s models can. This looks very much like confirmation bias on the surface, but the fact that it’s rational makes me think it’s not fair to characterize it this way.

Endorsing this too much gets a bit problematic, though, because it can easily lead to problems where you get too set in a single model while failing to realize other relevant factors. I think for this reason an adversarial approach should be used, where people on each side of the issue discuss the situation and try to find experiments that would distinguish the two theories. Then people could do the experiments that confirm their theory, and we would get progress on the issue. I’ve been trying to set this up, and it hasn’t worked great yet, but hopefully it will eventually.

An ending note: one aspect of my toy model that is a bit unfair is that I set P(E|I) = 0.05; many typology-skeptics also have serious reservations about the validity of PPG, so they would likely seriously discount the evidence obtained from this source. I don’t find their objections especially convincing, so I don’t think it’s fair for them to set P(E|I) very high, and that kinda puts us in a tricky situation. My interpretation of this issue is that they want to keep traits like sexual preferences difficult to observe directly so their politically-motivated theories on the subject can’t be empirically disproved. This is of course a cynical theory, but Aumann’s agreement theorem requires me to have cynical beliefs when there is persistent disagreement.