Age of onset as the origin of discrete types of gender dysphoria?

[Epistemic status: currently speculative, but probably easily testable when I get the time]

In a previous post, I created a model where different factors (GNC and AGP) influenced a single dimension of gender issues. This model yielded roughly speaking two types of trans women, one of which was homosexual and GNC, while the other was AGP. This split happened as a result of Berkson’s paradox, the phenomenon where if there are multiple causes that influence some common factor, and you select for said factor, the causes end up negatively correlated in the selected sample.

The Problem

However, my model was fundamentally flawed, because I needed to “fake” the parameters with unrealistic values in order to create the two clusters. In particular, I believe I set the rate of homosexuality higher than realistic, the rate of AGP among homosexual men too low (it’s probably more like 5%), the connection between AGP and gender dysphoria too weak, the degree of femininity among homosexuals too high, the influence of femininity on gender dysphoria too high, and the overall rate of gender dysphoria was likely also too high. This doesn’t really change the point in my previous blog post, since I get essentially the same connection between AGP transsexuality and femininity if I use more realistic parameters, but it does have implications for the theory of gender dysphoria.

Namely: we appear to observe two different kinds of gender dysphoria, and this dichotomy requires some theoretical explanation. Berkson’s paradox, which I relied on in the post to explain it, is too weak. It consists of the fact that selecting for the existence of some traits leads to a negative correlation between the traits, but it tends to more lead to a spectrum than to a dichotomy.


Illustration of attempting to achieve the trans typology through Berkson’s paradox. Each point represents a person, and the assumption is that people transition when their total gender issues – defined as the sum of the two etiological contributors – exceeds some threshold. While it will lead to a strong observable negative correlation, it will still be a spectrum, and both types of trans women will have elevated degrees of both etiologies.

If the etiologies were incompatible – that is, if the shape of the distribution above was an L shape, rather than a round O shape – then the center of the distribution would be removed, and we would end up with two discrete types. This would require that the etiologies were negatively correlated, and indeed negative correlations between the etiologies themselves (i.e. AGP ruling out GNC homosexuality) exist, but are also too weak (and indeed my model also included such negative correlations). So what are we to do?

There’s a different solution: What if the different types contribute to gender issues independently? That is, rather than adding the etiologies together to get the estimated degree of gender issues, we could take the maximum; this can end up cleanly yielding two distinct types:


In this modified situation, trans women of one type do not have any of the traits associated with trans women of the other type.

I’ve seen it argued that the psychological feelings arising from autogynephilia and HSTS-spectrum gender issues are fundamentally different, and that therefore this approach – taking the maximum – is more realistic, as it doesn’t really make sense to talk about “gender dysphoria” in general when there are multiple different kinds. However, I strongly disagree with this conclusion; both from theoretical concerns, and from just practical psychological concerns:

Consider, for instance, the process of “gender-questioning” that many autogynephilic transsexuals go through. In this process, one question that often comes up is the question of whether one is a “true transsexual” – and obviously things like gender nonconformity, childhood gender identity disorder, and so on, are all going to contribute to concluding that one is indeed a “true transsexual”, increasing the likelihood of transition. Thus, I think the additive model is far more realistic than the maximum-based model.

The Solution

So, what can we do? The title of course spoils this a bit, namely we can use the age of onset as the factor that gives us our “maximum”. Essentially, it seems realistic enough to assume (tautological, almost) that people ever develop gender dysphoria if they at some time develop gender dysphoria; so people’s life-total degree of gender issues can be understood as the maximum degree of gender issues that they have had at some point in their life. This gives us the maximum we have been looking for.

If we then further assume that different factors influence gender dysphoria differently at different life stages, we’re done! Specifically, if we assume that factors related to the HSTS spectrum of gender dysphoria (such as gender nonconformity) influence gender dysphoria more strongly earlier than later in life, while factors related to the AGP spectrum of gender dysphoria (such as, well, autogynephilia) influence gender dysphoria more strongly later than earlier in life, then this gives us something akin to the two types.


Required influences on gender dysphoria over time. Scare quotes around “HSTS” because HSTS is typically the term used to refer to the type of trans people, rather than the etiologies involved (which include e.g. gender nonconformity, but also other things).

The great thing about this assumption is that it is already broadly accepted as true. Specifically, it’s generally accepted that the vast majority of kids with gender identity disorder desist, indicating that whatever contributed to their gender dissatisfaction in childhood could not continue to affect them later in adulthood (and conversely, it’s generally accepted that HSTSs were GID kids when they were younger); and it’s of course also often accepted that AGP has a larger influence on gender satisfaction from adolescence and onwards, when libido becomes stronger.

The HSTS cluster of gender issues probably consists of multiple effects, and to have a model that encompasses all nuances in play, these multiple effects should be analysed separately. Any such analysis will currently be speculative, but I can give an example of what I mean:


Possibility for the change in contribution from different factors associated with HSTS. Here, cognitive gender confusion refers to a belief that one is female; attachment/ingroup factors refers to having a bad relationship with one’s father, or similar inverted with mother, or to idealizing girls and having a negative view of boys, gender norm enforcement refers to being bullied or otherwise harmed because one is targetted by other individuals for being gender nonconforming, and desire for masculine heterosexual men refers to the HSTS attraction to, well, masculine heterosexual men, that Bailey described in The Man Who Would Be Queen.

Most likely, the AGPTS cluster of gender issues could be similarly decomposed, but it’s less clear how to even begin with that. Autogynephilia itself does not seem to be the only thing in the cluster that contributes to gender issues, as e.g. bisexuality correlates with autogynephilia and appears to contribute independently of autogynephilia (at perhaps d~0.75). However, it seems implausible that bisexuality is the direct contributor, as there’s no clear reason why it would be, so it would need to be broken down further to be interpretable.


The observations above do not make the homosexual/autogynephilic typology obsolete. Rather, they suggest that there are two ways of viewing things. There are the etiologies of gender issues – various groups of traits that correlate internally and produce gender dissatisfaction. And then there are the onset clusters, which determine what sets of gender issues are able to get combined together, timing-wise, to lead to transsexuality.


The two different perspectives, etiology vs onset, determine whether you are focused on what traits co-occur and influence each other, or whether you are focused on what traits trans people can end up having.

One perhaps-intuitive way to think about this is to consider the question of “building a trans woman”. Most natal males don’t transition, so in order to end up transitioning, they need some combination of unusual traits. No single trait can predict transitioning, so therefore it is insufficient to have only one trait, and we instead need to consider some combination of traits that are sufficient. If we abuse the liability-threshold model a bit, we can even come up with a scoring system; different traits contribute differently to one’s scores, and they also contribute differently depending on the onset age. In order to end up gender dysphoric, the set of traits must add up to a sufficient threshold of gender issues – which, based on the transition rate, might be perhaps 4 sigma (though perhaps it’s lower than that, say 2.6 sigma…).


Different contributors to gender issues, visualized as arrows in proportion to their effect. To determine the numerical score of different contributors, I consider the correlation between the contributors and gender dissatisfaction.

The more contributors one adds, and the more extreme the contributors one adds, the more “unlikely” the individual one is writing down is. To illustrate the system, here is some sort of attempt at scoring myself using this model:


Polycausal overview of factors that plausibly influenced my gender dissatisfaction. To be honest, I’m slightly surprised that the model was so good at accounting for my gender issues.

When I was younger, the AGP had much less effect, and it’s also likely that the masculinity had much greater effect. This matches with me having developed gender issues late. However, if someone had been more feminine than me, then they might have started out with some milder gender issues, which then dissipated as the effect of femininity got smaller with time, but also increased as the effect of AGP grew. Thus, this model predicts that there can be some quite-complex patterns of evolution of gender issues over time.

Generally, as one adds more and more contributors in the model of an individual, the likelihood of all of these contributors existing in the same person drops. This is the main factor that puts order into the ways gender dysphoria can function; an individual can only have so many contributors to gender issues, so it is unlikely for them to develop gender dysphoria if they don’t have the big ones (such as autogynephilia for late-onset individuals).

Some additional complexities

The model currently assumes that if one ever exceeds the (4 sigma?) threshold of gender dissatisfaction, one becomes permanently transgender. This assumption is a bit unrealistic, and is directly contradicted by phenomena like desistance.

It seems to me that there is some sort of nonlinear effect that leads to similar consequences to this assumption, though the specifics of the effect seems to possibly differ by age. For instance, it’s often believed that it is impossible to treat full-blown gender dysphoria in adulthood; this might be due to some sort of commitment or self-reinforcing effect, where gender issues strengthen once they exceed the threshold. Meanwhile, in childhood, gender issues seem to affect one in different ways, e.g. leading to pretending to be the opposite sex in ways that plausibly contribute to gender issues in the future.

I have some more thoughts on the details of this which I might address in a later blog post, but if there’s any flaw in the model, this is a likely where it is.

It is also worth noting that some trans women may have had mild gender issues in childhood that were not enough to lead to full-blown gender identity disorder, and then later in life developed gender issues for different reasons. For the purposes of this model, the milder earlier gender issues are not very important, as they represent gender dissatisfaction due to different causes than the ones that later made them trans.


Right now, all of this is theoretical speculation, so the primary implication is probably that this should be tested. And to the degree it is tested, it may also be useful to start collecting data on differences in how traits affect childhood versus adulthood gender dissatisfaction.

The model also suggests some ways in which people can vary from the standard HSTS/AGPTS dichotomy. These probably aren’t going to massively reorganize things, but they might make the typology more nuanced. At this point I’ve already started reinterpreting things in the light of this model, and it seems quite promising.

It seems to me that this model might also be better able to account for trans men. In particular, women seem to face stronger gender norms, which implies a larger effect size for gender nonconformity on adulthood gender issues, and thus a greater degree of “blurring” the types.

Economics of transition, the liability-threshold model, and gender dysphoria

I recently wrote a blog post with a model of how people end up becoming trans, and in response I received an email asking about the foundation of the common “gender dysphoria” variable that all the different causes contributed to. I thought I would write a blog post about what the meaning of this is:

In economics, the simplest model we use to understand people’s behavior is the utility-maximization model. Under this model, people have a set of preferences, which is defined by their utility function, which assigns outcomes a number that represents the desirability of said outcomes. When people have a choice, they are then assumed to pick the option which leads to the highest expected utility.

We can apply this model to transgender topics. The core defining property of transgender people is that they transition, and ultimately transition is a decision to be made. If we apply the economic model, then we get the conclusion that people become trans when utility(transition) > utility(staying cis).

These utility functions are relatively complicated objects, especially because for our purposes, they represent expected utility, which means that they also include an element of belief about what is going to happen in each situation, and not just what is actually going to happen. Thus, utility(transition) is likely going to contain terms related to autogynephilia and passability, but also things related to whether one believes that transition is a good option for trans people in general. They also don’t just include one’s own psychological traits; for instance, utility(staying cis) might be higher if one has a successful established life (significant other, job, …) as one’s natal sex.

Thus, we might imagine that we can approximate utility(transition) = autogynephilia + passability + femininity – transphobic environment + …, and that one can approximate utility(staying cis) = masculinity + relationships + attractiveness as natal sex + …; with the different factors influencing one’s decision being included in the model.

The specifics of how these preferences work (e.g. whether they are based in pain or yearning, etc.) don’t matter from an economic point of view. There is perhaps a sense in which one can say that utility(trans) is likely to represent “positive emotions” related to transition, while utility(cis) is likely to represent “negative feelings” about being one’s assigned sex, but ultimately they’re treated symmetrically, and so the distinction is not very important. This may make it seem like the economic understanding is missing critical information, but depending on the purpose, this may be perfectly acceptable.

In the liability-threshold model used in behavioral genetics as well as other fields, it is assumed that for some binary condition (such as being trans), there exists a latent “liability” to develop this condition, such that when the liability exceeds some threshold, then one ends up with the condition. However, notice what happens when we combine this with the economic model above: utility(trans) > utility(cis), which is the condition for ending up trans, is equivalent to 0 < utility(trans) – utility(cis). This gives us a rather complicated variable, utility(trans) – utility(cis), which leads to transsexuality when it exceeds a certain threshold – exactly what the liability-threshold model needs! So, we can define liability(trans) = utility(trans) – utility(cis). In reality, this is going to be a bit more complex than what I’ve described above, as there may be nonlinearities, change across time, irrationality, etc., but it’s a good starting point.

Another thing to notice is that this notion of liability(trans) is very close to how gender dysphoria can end up defined in clinical settings. For instance, in the factor analysis of the GIDYQ-AA, items related to intention to transition, such as wanting HRT or SRS, have incredibly high factor loadings at around 0.95. From a psychological standpoint, it might be reasonable to try to figure out some narrower sense, which can apply to someone who decides not to transition despite discomfort, and which might not apply to someone who wants to transition while having relatively little discomfort, but there’s also a real sense in which the decision-oriented approach makes sense.