Reflections on a failed attempt at testing the causal relationship between autogynephilia and gender issues

I’m really deep off a tangent, so I’m not sure this post will be understandable to anyone but me. 😅

I am as always trying to understand the causes of affective gender identity, and for this in males there is one thing that will always remain striking: autogynephilia, i.e. a sexual interest in being female, is highly correlated with wanting to be a woman, to the point of being one of the factors that is most strongly associated with gender issues.

And this is, as always, associated with the big question: What’s the causal relationship here? The correlation is pretty irrelevant if autogynephilia is simply a symptom of wanting to be a woman, as the goal is to understand the causes rather than effects of gender issues. But how do we know?

As a Blanchardian, I believe autogynephilia causes wanting to be a woman. An argument Blanchardians often make is that we can observe this from autogynephilia preceding gender dysphoria, but I think this is a terrible argument, for two reasons:

  1. While autogynephilia might usually precede full-blown gender dysphoria, it appears simply wrong that overt autogynephilic arousal always precedes all of the cross-gender ideation it is associated with. Blanchardians often explain this via proto-sexual ideations, analogous to childhood crushes; and I accept that explanation, but this implies that there are dynamics in play before the temporal events that the argument relies on, and therefore the argument is fundamentally flawed.
  2. More generally, you can’t in general know that such events wouldn’t be there; observing two correlated events separated in time is not enough to tell you that the first causes the second, as there might instead be confounding.

So if this argument is flawed, why do I believe autogynephilia causes gender issues? Ultimately I think it has to come down to: We know what sexuality is, what it does. The entire point of your sexuality is to motivate you; it’s not particularly plausible that it is caused by or confounded with gender issues, compared to it causing gender issues.

As in, suppose you forget everything we know about sexuality and wanting things. Instead, we just observe a correlation between Thing A associated with women, and Thing B associated with women. Even if Thing A comes before Thing B, is there much reason to suppose that Thing A causes Thing B? No! Of course not! Correlation != causation. But conversely, if we forget everything we know about women, and just observe a correlation between a sexual interest in Thing C and a desire for Thing C, is it then reasonable to think that the sexual interest causes the desire? Absolutely! Because that’s what sexuality does.

… except, how do I know that? For this, I’d bring up several arguments. It makes obvious sense evolutionarily speaking. It’s how we usually think of sexual interests like attraction to men or attraction to women. It’s not like we find it easy to control our sexuality – indeed, autogynephiles who dislike being autogynephilic and who don’t want to be women would seem to contradict the notion that it’s merely a question of sexuality reflecting desires – but this is fully compatible with a sexual model, as they resemble ego-dystonic gay men.

But sexuality being this deeply malleable thing that reflects our deepest desires does seem to reflect how critics of Blanchardianism see things. In fact, it seems rather central to their critique. Sometimes it can even look pretty plausible to me – wouldn’t we expect a gender dysphoric male to imagine being female in ordinary sexual fantasies? And maybe autogynephilia is somehow different from other sexual interests? Or maybe the understanding of other sexual interests is mistaken?

I believe I have come up with an approach for investigating these sorts of questions, and I ended up getting impatient and testing this approach, but it turned out not to work. I don’t think the failure is fundamental to the approach, but instead due to the low quality of measurement I did, so I think it would be worth talking about the method.

Psychological systems

How can I even talk about “what sexuality does”? I’m taking things that happen for some sexual interests, such as gynephilia, and generalizing them to entirely different sexual interests, such as autogynephilia. How can that make any sort of sense?

It makes sense only if we suppose that they both represent variations in some common underlying system – in this case, sexual preferences. That is, I suppose there must be some innate system that I call “sexuality”, which contains certain preferences and motivates people to act on them. If I didn’t believe that – if I believed that sexuality was just a mishmash of different things lumped together under the same label – then it would make no sense to believe that I could generalize from gynephilia (or other sexual interests) to autogynephilia.

So let’s consider this sexual system more specifically. I believe that it contains some latent sexual preference. I further believe that upon encountering instances or fantasies of this sexual preference, one tends to experience genital arousal (though this is moderated by contextual factors like state libido) in proportion to its fit to the preference. I furthermore believe that this preference somehow leads to a motivation, a sort of desire, to achieve one’s preference.

I think this is simultaneously rather basic, carricatural, and obvious. But it gives us a basic foundation: for the purpose of measurement, we might equate the sexual preference with the arousal pattern. We then model that propensity for sexual arousal to something causes general desire for this thing, but that the general desires might be affected by other factors to.

In order to capture that there is a single system underlying all of this, it might perhaps be reasonable to propose that there is a single coupling constant k, such that:

desire for x = k * arousal to x + other factors

It turns out that this hypothesis makes some very strong predictions, much stronger than you usually see in psychological research. We can get into this mathematically.

(I guess feel free to skip the next bit if you’re not strong on math… 😅)

Specifically, suppose that arousal to x has a variance of A. Then desire for x ends up with a variance of k2A from the arousal, as there will be differences in wanting x depending on how arousing one finds out. This will also imply a covariance of kA between arousal and desire. However, in addition to this, desire for x may have some additional variance B’ due to other factors not related to arousal. (E.g. homophobia may reduce interest in same-sex partners, extraversion might increase desire for partners, femininity might increase desire to be a woman, and so on.) So the total variance of desire for x, which we will label B, can be written as B = k2A+B’. Using the variances and the covariance, we can then compute the correlation: r = kA/√(AB) = k√(A/B).

Why are point predictions important?

The expression that I derived for the correlation above predicts the exact strength of the correlation between a sexual interest and its corresponding desire. (Sort of – it makes some unrealistic assumptions that would need to be addressed to get anywhere close to exact results.) This is unusual in psychology; typically, psychological research involves grabbing some folk commonsense guess at how psychology works, making directional (i.e. positive/negative) predictions about some correlations, and then testing those directional predictions.

There’s a lot of problems with this. First, directional predictions are generally just going to be right half the time; there’s only two directions the association can be, so you can only do weak hypothesis tests with them. Furthermore, because there’s generally content overlap between the things one makes predictions on, the probability that there is a directional effect is going to be even greater. Thus, predicting the strength of the correlation is a much stronger hypothesis test.

But even more importantly: the equation I derived for the predictions treats the two directions of causality differently. The correlation increases with the variance in the cause, and decreases with the variance in the consequence; thus, assuming that one has multiple parallel systems, point predictions allow testing the direction of causality.

Now, the equation I derived this for assumes a rather simple linear effect. Maybe (almost certainly…) sexuality is more complicated than this, and so the equation won’t hold in practice. But if one studies sexuality more carefully, maybe one can find some model that does hold more precisely. At least, if one can’t, then that casts doubt on the assumption that sexuality forms a consistent system that one can generalize across. (Almost by definition.)

Autophantasic sexuality

So basically, an approach could work as follows: take some other sexual interest with different variances than autogynephilia, use this sexual interest to fit the parameter k, then apply k to the case of autogynephilia and check whether it gives the right point prediction. In practice, this is a bit optimistic; we’d probably want to run many sexual interests in parallel to better check the robustness and validity of the theory asserting that sexual interests are a real thing, and we probably wouldn’t expect an exact fit just because the assumptions made are a bit optimistic. But this is the general approach.

In practice, it gets a bit more complicated than that. What corresponds to what? For instance, when asking about sexual attraction to women, should one consider arousal to the nude female body, arousal to having sex with a woman, or something else? And when one asks for desire for women, should that be desire for a girlfriend (and if so, what defines “girlfriend”? just everything the local culture includes?), desire for sex with a woman, or something else? How does this generalize to submissive bondage, should desire for bondage refer to desire to be tied up, or desire to have a master who will tie oneself up, or what?

These are the sorts of questions that research into the construct of sexual interests should start tackling, in order to increase understanding. But for now, my solution is simple: restrict the investigations to the narrower case of “autophantasic sexuality”; i.e. sexual interests in being something specific. This “being” something can still be broad; it covers e.g. furries, ageplay, and much more. So:

Definition: If X is any trait that one can have or any category one can belong to, then we consider the propensity to get sexually aroused by the thought of being X to be an autophantasic sexuality, which we label autoXphilia. There is some coefficient k independent of X such that autoXphilia causes a desire to be X.

Autophantasic sexuality resembles the concept of an “erotic target location error” (a sexual interest that has been inverted onto oneself), but unlike the case of ETLE, it does not require attraction to the original target.

It’s sort of premature to test it, as one would need a good measure of wanting to be X as well as a good measure of sexual arousal to being X. But my simulations suggested that this wouldn’t be overly sensitive to violations of assumptions/bad data quality/etc., so I went ahead anyway. To test this model, I collected a broad range of Xs:

  • Woman
  • Child
  • Furry (described as “an anthropomorphic animal (i.e. keeping a roughly human form, but having fur and animal features)” to participants)
  • Amputee
  • Extremely fat
  • Asian
  • Black
  • Androgynous
  • Exaggeratedly muscular
  • More attractive
  • Extremely rich
  • Nudist (“living nude in your everyday life, in a society of other nudists”)
  • War hero
  • Famous left-wing activist
  • Manic (“extremely excited, hyperactive, prone to grandiosity, distractable, impulsive…”)
  • Android (“fully-functional human-shaped robot”)
  • Extremely skinny

There’s some comments that I should make on this list. My assumption would be that some of these sexual interests would not exist; for instance I have never heard of auto[left-wing activist]philia. However, it is intentional that I have included these. If my assumption is right that the sexual interests do not exist, then the variance in the sexual interest would be ~0, which would mean the correlation between the sexual interest and the desire to be a member of the category would be ~0. (In practice I fit using the covariance rather than the correlation to avoid worries about dividing by zero.) On the other hand, if my assumptions are wrong, then including counterexamples like these help prove that the assumptions are wrong.

Another important thing is to consider interests with different properties. For instance, my assumption going into this was that a primary factor in wanting to be an amputee would be sexual; after all, why else would one want it? So my assumption was that for being an amputee, there would be high variance in the sexual interest, and lowish variance in the general desire. Meanwhile, for something like being a left-wing activist, I assumed the opposite pattern; how appealing this would be would be heavily dependent on how left-wing one would be, for instance.

These are a lot of assumptions, which brings a third important thing. I don’t know which of the assumptions are true or not before I collect the data (and sometimes not even after it); that’s why they’re assumptions rather than proven facts. Thus, it is important to study a large number of potential interests at the same time, in order to achieve robustness against violations of assumptions.

Finally, what was my plan for measurement: Simple, for each of the possible things one could be, ask people how appealing they would find it on a 5-point scale from very unappealing to very appealing, as well as how sexually arousing they would find it to imagine being this on a 5-point scale from not at all to very. This is a really terrible measurement; it is discrete and contains a ceiling as well as a floor. Thus, another important check against robustness is to include targets that bump against the limits of the measurement; for instance I would assume ~everyone wants to be rich, so it seems conceivable that being rich would bump against the ceiling of the measurement, which violates the assumptions made in the calculation. In order to find out how serious that sort of violation is, I have included extremes like this.

Initial results

I did the first round of this survey on Prolific a few weeks ago. The results were not very convincing. To explain, let me introduce a strange kind of diagram:

Unusual diagram I came up with for this sort of hypothesis testing/causal inference.

The equation r=k√(A/B) predicts a relationship between the variance ratio of autophiilia/desire to be, and the correlation between the two variables. As such, to get an overview of the results, one can do a scatterplot, showing how each of the autophantasic sexualities fit in. If the theory is right, they should all be on a line. I wanted to plot this, but A/B is unbounded, so to make it better behaved, I instead plotted A/(A+B), which is essentially similar, except that it ranges from 0 to 1 and thus more neatly fits into a plot. The consequence of this is that the interests would be expected to lie on nonlinear curves, rather than straight lines.

I’ve fitted two constants, one for autophilia causing desire, and the other for desire causing autophilia, and plotted the curves associated with these constants; the orange curve represents autophilia causing desire, while the green curve represents desire causing autophilia. As can easily be seen, they both make very distinct predictions, but they are also both very very wrong.

More specifically, it appears that the interests lie on a sort of upside-down U shape; it starts at the bottom-left with attractive/rich, continues up to the top-middle at fat/muscular, and then goes down to the bottom-right with left-wing activist/child. Meanwhile, the autophantasic sexuality model would predict that you just see a decrease, which fits badly with attractive/rich, while the reverse causality would predict that you just see an increase, which fits badly with left-wing activist/child.

I see this plot as being a good illustration of my point that I needed a broad set of interests to test my assumptions. Some assumption were wrong; e.g. the position of amputee and left-wing activist wasn’t much different, contrary to what I predicted. And for seeing the shape of the curve, it helped that I had a broad set of items; if I had e.g. excluded being rich from the test, it would be unclear if “attractive” was just an outlier.

One thing I have previously found by experimenting on Prolific is that the results are not as consistent as I would like. Just because people say one thing one time does not mean they will say the same thing another time. So to see what fraction is due to the persistent aspect of people’s responses, I did another survey with the same participants one week later to extract only the component of their responses that was consistent across surveys. That yielded the following plot:

Correlations and variances after accounting for the inconsistent fraction of participants’ responses.

The main thing this changed was the effect sizes involved. We still have the general upside-down U shape, and the rough placements of different items is still the same. Clearly the data doesn’t support the simplified approach I’ve taken.

Ordinal data, interval data, ceiling effects

I believe the biggest problem is in measurement. Everyone wants to be richer and more attractive. As a result, the two options, “appealing” and “very appealing”, are not very good for distinguishing people; whether one thinks being rich would be super great but not worth the effort, or whether one has it as one’s biggest goal in life to be rich, is not something that this measurement distinguishes; they would both go under “very appealing”.

One can talk about these problems more formally. The trouble is that I am working with “ordinal data”; I have an ordered set of categories that people can respond that they belong in, but distances between the categories, or widths of the categories, are not really well-defined. Thus, it is mathematically incoherent for me to talk about the “variance” in people’s responses.

In fact, how do I compute this variance? The standard way: I assign each response option an integer from -2 to 2, and then compute it using these numbers. That is not mathematically valid. (People do it all the time in psychological research, but usually their methods are not as sensitive to the variances of the variables as my method is.)

What I need is “interval data“; data where concepts like variance and relative differences are mathematically meaningful. Otherwise, I will run into “ceiling effects”, where highly varied responses might be “squished” together into a small variance in the quantitative data. You can think of temperature as being an example of interval data; differences in temperature are quantitatively meaningful, such that it makes sense to talk about a difference of 10 degrees Celsius. (Meanwhile you can’t talk about differences on a scale from “very unappealing” to “very appealing”; it’s not clear that the distinction between “unappealing” and “neutral” is the same as the distinction between “appealing” and “very appealing”.)

So can one do that? Measure preferences in interval data? What does that even mean, when preferences are mainly defined by an ordering (inherently ordinal!) of what one wants?

One possible answer to this comes with rational decision theory; it turns out that someone who deals rationally with uncertainty can be proven to have preferences defined in intervals rather than just ordinally, because they must be able to take weighted averages when dealing with uncertainty.

Thus, if one makes the assumption that humans are rational, one can extract their preferences by hearing what tradeoffs they would choose if there were possible random chance involved. Roughly speaking, if one has some worst scenario D (e.g. death), and some best scenario B (e.g. being rich, popular, attractive, etc.), then people’s preference for any scenario Q can be defined as the probability p such that they are indifferent between having Q, and having p probability of B and (1-p) probability of D. This does run into the problem that humans are not fully rational, and one would probably need to find some way of dealing with this.

A bigger problem is that while this allows intrapersonal preference comparisons, it does not allow comparing one person’s preferences with another’s, which is necessary for computing correlations and variances across people. Specifically, if people vary in how much they value the best scenario relative to how much they value their sexuality, then comparing relative to their best scenario would be a problem. It seems like it should be possible to handle this, though; rather than standardizing people’s preference measurements in terms of the best and worst imaginable scenarios, one could standardize them in terms of points that are relevant for this research, like the importance of having a sexual relationship.

The main problem I see here is that measuring preferences properly seems pretty expensive. The badly measured study I did on Prolific was already quite a bit more expensive than any research I’ve done before, and proper measurement would presumably make it much much more expensive.

But I think it would be worth it! If these sorts of methods can genuinely settle controversies like autogynephilia (as well as, IMO, many other psychological questions), then the expense may be worth it, at least compared to other methods that mainly symbolically test things.

Is there a need for a psychophysics of autogynephilia?

The main measurement problem I worry about is in measuring preferences. But a secondary one I worry about is in measuring and defining autophantasic sexuality. More specifically, I got quite high endorsement rates to some sexual interests I did not expect to get high endorsement rates for, such as being rich.

Maybe this is just common. Or maybe it’s uncommon, but common on Prolific (after all, autogynephilia is also fairly common on Prolific). But due to some other questions I have been interested in (such as defining autogynephilia in such a way that it is validly measurable in women), there’s a pathway of investigation that seems worthwhile, namely: What does it mean to be aroused “by” something?

I find it hard to communicate the issue. It’s not a question about causality per se; I’m fine with taking causality as a concept for granted. I think we can meaningfully consider different sexual scenarios and define how arousing each of them is. The issue comes to when we want to abstract the arousal to be defined by features of these scenarios, rather than simply considering specific scenarios.

I think there are a number of paradoxes that can be raised. For instance, if someone is attracted to people regardless of their sex, we say that they are bisexual and aroused by both men and women. However, if someone is attracted to people regardless of their hair color, we don’t say that they are panhairsexual and aroused by all hair colors. Is that a purely linguistic paradox? I don’t know, it doesn’t feel like it, but I’ve thought of various solutions and they didn’t seem convincing.

One can also pick a paradox more specific to AGP; if a man is aroused by the thought of engaging in sexual activities as a woman, then we say that he is AGP, even though we don’t say that he is AAP if he is aroused by the thought of engaging in sexual activities as a man. One might argue that this categorization is overly broad, and that being a woman must be arousing independently of the other things for it to be true AGP; however, this does not seem viable to me, as in surveys generally men who report being aroused by the combination of being a woman with other things also report being aroused by the thought of being a woman on its own; any viable theory of AGP needs to account for this connection.

I don’t think I have a full account for how to interpret this. However, I don’t think we need to assert one. Rather, the appropriate method is to study it empirically. Ideally we should be able to examine people’s arousal to a variety of concrete scenarios, and abstract them into a smaller number of parameters, such that those parameters well match their actual interest in the scenarios. This should ideally be done very precisely, with point predictions, as this allows more specific test of differing theories. This sort of research is called psychophysics, and my understanding is that in psychophysics, the issues related to ordinal scales that I mentioned before aren’t as problematic.

If the exact relationships for AGP were understood, one could also use this to better delineate and measure other forms of autophantasic sexuality. One could meaningfully restrict oneself to only the forms that match AGP in shape, and one could perhaps also get a better understanding of what it is that one is measuring.

Conclusion

Overall, I think autogynephilia research could benefit from a greater focus on foundations. In fact, this is probably not just limited to autogynephilia research; all psychological research needs a better understanding of the low-level foundations. The big problem is that it is expensive and slow, and the measurement is difficult. However, it seems like the only way to make solid progress.

There are alternative ways to make progress. Behavioral genetics manages to do a lot with twin studies, sometimes one can use instrumental variables, and so on, but these methods are generally extremely specialized and fiddly, and usually disregarded as a result. Meanwhile, zooming in to the narrower, more specific case allows solid progress to be made, as long as one actually starts from a valid concept.

That’s not to say that there isn’t space for more usual kinds of research. However, they should really be seen as exploratory research that does basic sanity-checks of hypotheses, rather than the confirmatory research that a lot of studies present it as being.