“While both of Ngo’s claims are technically true, the proper context paints a widely different picture.”
In late November of this year, controversial media figure Andy Ngo was suspended from Twitter. As we are all frequently reminded, Twitter is not real life. This is certainly true, and generally I do not find the discussions of events that happen on Twitter to be all that interesting. This, however, is a particular case, because I think it is better described as an example of common argumentation tactic that happened to occur on Twitter. According to The Post Millennial, for which Ngo serves as editor-at-large, he was unfairly suspended for stating nothing more than mere objective facts. The specific incident that got him suspended was a reply to Chelsea Clinton in which he pointed out two facts. First, that the homicide rate for the transgender population is lower than that of the general population; and second, that the majority of those homicides are committed by black men. The article mentioned earlier from The Post Millennial reports that after this tweet, Ngo’s account was suspended for violating guidelines on hateful content. The reaction that followed, in particular from right-leaning figures, like those on the Intellectual Dark Web (or adjacent personalities) is a clear example of an insidious phenomenon that I believe can only be described as either statistical illiteracy—or bad faith. As I will explain further, I think the latter is the more likely conclusion.
The reaction from right-leaning and IDW-adjacent circles was nothing if not predictable. Public intellectuals like Christina Hoff Sommers and Peter Boghossian quickly denounced the suspension as a politically correct action by Twitter, targeting Ngo for stating uncomfortable facts. Both of them are accomplished academics in their fields, and while they may not be statisticians or work in fields specifically dependent on statistics, the statistical nuances behind Ngo’s claim are not so complicated that only someone with advanced knowledge in the field could comprehend them. It is for this reason that I believe that when they claim foul play from platforms like Twitter and play innocent regarding the nature of a technically factual claim like Ngo’s, I have to believe that the reason is bad faith rather than plain lack of knowledge of statistics. And, after all, it is no secret that statistical facts can be used to mislead and lie.
Let us assume for a moment, however, that what is behind this is actually lack of knowledge, and, from there, I will explain what is going on. Both of Ngo’s original claims are technically factual, but they are highly statistically misleading. A 2017 study on transgender homicide rates published in the American Journal of Public Health acknowledges that the first part of the tweet is true. The paper examines known homicides from 2010 to 2014. Using an analysis of known risk factors for homicide, the author concludes that even controlling for underreporting, risk factors of the transgender population predict a lower homicide rate than that of the cisgender population. The problem, however, is that the global homicide rate has very little use as an analytical tool. The same study also explains that when broken down by gender, transgender women have a higher homicide rate than the overall rate for women. Despite opposition to the use of specific gendered pronouns and so-called ‘gender ideology’ in general, it really makes more sense from a sociological point of view to compare homicide rates of transgender women with those of cisgender women. Trans individuals assume the social roles of their gender identity, not that of their biological sex, so, in terms of risk factors, they should be treated as their preferred gender. Doing the opposite would only serve a political agenda. And once this is done, the claim that the trans murder rate is lower than that of the overall population reveals itself to be meaningless. This is because once the rates for women were taken on its own, it became apparent that being trans does pose an added risk for becoming a homicide victim. If it did not, the trans rate would be similar to (or lower than) that of cis women.
Now, it should be stressed that everything said so far should be taken with caution and maybe even skepticism. A different 2017 paper in the same journal analyzes the quality of the data on transgender homicides and concludes that the data has significant limitations. These limitations include everything from the actual total of transgender homicides, to the real size of the transgender population. In that sense, it is not possible for me to conclude unequivocally that transgender women have a higher homicide rate than women overall. The currently available information does suggest that Ngo’s claims are not entirely accurate. What can be concluded with certainty, however, is that the use of statistical facts is at best careless and at worst intentionally misleading. As we will see in what follows, this is a recurring theme.
Let us now look at the second claim: that the only fact stated is that the majority of homicides with trans victims are committed by black men. Note that the actual percentage of perpetrators is never mentioned. This omission alone should be a cause for concern and serve as reason to doubt the motivations for making the claim in the first place. To illustrate why, let us use a hypothetical example. Assume that 60% of vinyl record purchases are done by men and the remaining 40% by women. Now, vinyl records represent a very small share of how people get music. In all likelihood, a majority of people get their music from online services. If we were to assume that the distribution of men and women’s taste for different genres is equal, then we would observe that when we break down vinyl sales by genre, the 60-40 distribution of purchases remains constant. In this scenario, if someone were to claim that men like jazz more than women because 60% of all jazz vinyls are purchased by men, it would show that this person has a limited understanding of statistics. This breakdown is already expected—because overall men buy more vinyls. What would be noteworthy, on the other hand, would be if we observed that men buy 75% of all jazz vinyls, but women buy 55% of blues vinyls. In both of these cases, there is a deviation from the expected outcome. In this case, it might be more likely for us to conclude that men like jazz more than women—and that women like blues more than men.
But as I just said, the fact that a majority of trans murder victims are killed by black men means nothing in itself because that is already the case for homicides in general.
To get back to the real case at hand, in the United States, it is already known that, overall, the majority of homicide offenders are black men. Explaining the structural and sociological factors that might contribute to that figure is well beyond the scope of this discussion. For now, then, it will be simply taken as a starting point. The FBI Unified Crime Report, which gives disaggregated data regarding homicide victims and offenders, states that, from 2015 to 2018, 24,477 of all homicide perpetrators whose race was known were black. Of the known remainder, 19,649 were white, and 1,200 were listed as “other.” This puts the percentage of black perpetrators at 54%. Let us back up for a moment. The original point which Ngo was trying to refute was that transgender people suffer violence disproportionately. We already saw why his attempted rebuttal is flawed. But leaving that aside, the race of the offenders is completely irrelevant to his argument, so pointing it out is already suspect. This should not be controversial, which is why I find the feigned ignorance about what was wrong with the tweet to be, at the very least, annoying. I cannot see what could be added to the discussion by noting this last fact, other than implying that the black community has a problem with transphobia. But as I just said, the fact that a majority of trans murder victims are killed by black men means nothing in itself because that is already the case for homicides in general. If one wanted to make that case, what would need to be shown is that the percentage of trans homicide offenders which are black men significantly differs from the baseline, but Ngo did not bother to do that, which ought to make one question his motives.
As I have explained before, there are major problems with the data regarding homicides with trans victims. The issue of offenders makes this even worse because a significant percentage of perpetrators are not even known. When comparing percentages, it is not enough to know that they are different. When working with limited data, every data point is always given with a certain precision level—more commonly known as margin of error. Moreover, that margin of error is only given for a particular confidence level. In the simplest terms possible, this could be understood as the probability that the real value that is estimated by the statistical analysis actually falls within the given margin of error. So, for example, if a sample of blues vinyl buyers finds that 55% of them are women and 45% are men, that is still not the full picture. Given the sample size, the correct way to give the information would be that the proportion of blues vinyl buyers that are women is 55%, with a precision level of 5% and a confidence level of 95%. So, the interpretation of the full estimated statistic is that there is a 95% chance that the percentage of blues vinyl buyers who are women fall somewhere between 50 and 60%. The desired sample size when estimating proportions depends on three factors: the precision level, the desired precision level, denoted by d; the Z-score, which corresponds to the desired confidence level, and the expected proportion, denoted by p, which is usually taken from previous studies—or assumed to be .50 for maximum variation, which gives the maximum sample size. The Z-score for a 95% confidence level, which is usually the lowest acceptable one is 1.96 The sample size is given by the following formula:
Sample size = Z^2 × p(1-p) / d^2
Plugging in values for a 95% confidence level and a precision level of 10% (assuming maximum variation), the sample size would be 96. A precision level of 10% is already rather broad, and 95% is the minimum acceptable level, so a sample size of 96 is already weak, though it can allow for some conclusions as long as the estimated proportions are large enough. If we return to the vinyl example that I just gave but using the statistical parameters for a sample size of 96, the margins of error for each estimated value would be 55%±10% and 45%±10%. If this were the case, the margins of error would overlap, which would force us to conclude that the difference is not statistically significant. Given the margins, it would even be possible that the real value for men buying blues albums is larger than that of women, but the fact that the sample gave a lower one is simply due to chance. In the case of trans homicide perpetrators, we could use the expected 54%, rather than the 50% used for maximum variation since we have that data from the FBI Unified Crime Report. However, using that value only lowers the required sample size from 95 to 95. This is a major problem for anyone trying to make Ngo’s case.
An article that appeared in The Federalist, which tries to make a case similar to Ngo’s (and is cited by the original article in The Post Millennial, which denounced Ngo’s alleged unfair treatment), cites some trans homicide statistics. Specifically, it cites the 118 homicides between 2015 and 2019, as reported by the LGBT advocacy organization Human Rights Campaign. Of these, only 52 have known perpetrators. The FBI Unified Crime Report for 2019 has not yet been published, which is why I cited the data for the period between 2015 and 2018, since it is the closest. Before delving deeper into the specific data for trans homicides, further considerations must be made. First, the race of the victim and the offender are very highly correlated. The FBI data for the previously cited years shows that for murders with white victims, the perpetrator was white in 82% of cases, whereas for black victims, the percentage rises to 91%. The second important fact to keep in mind is that the percentages of victims for transgender people by race does not match the percentage for the total population. In the latter case, when the race of both victim and offender is known, 52% of victims are white and 44% black. Under the same circumstances, for transgender individuals 71% of victims are black, which already implies a huge difference in racial dynamics. Finally, as Human Rights Campaign asserts (a point that the piece in The Federalist echoes) that not all of these homicides can be attributed to transphobia. The HRC website does explain that it is still relevant to know when trans victims are involved because they are much more likely to be in situations of increased risk of homicide, such as poverty, addiction, and sex work. However, only four cases were definitively linked to hate, meaning most of those homicides can be attributed to the increased risks to which the trans population is exposed to. This is consistent with the findings of one of the studies cited earlier which found that—all else being equal—being a member of the trans community does pose an additional risk of becoming a homicide victim, especially for women. This becomes even more stark if we consider that for the overall population, only around one third of homicide victims are women, according to FBI data, whereas in the case of trans persons, a majority of the victims are women. Again, sociologically, trans women and trans men, as a general rule, fulfill the social roles of their preferred gender, so comparing trans women to cisgender women is the intellectually correct way to do the analysis—and has nothing to do with political correctness.
While provocateurs like Ngo will always be around, I do think it is time that public intellectuals on the right-of-center be held to much higher standards, especially given that many of them like to present themselves as dispassionate analysts and purveyors of facts.
We already know that the sample size we have of trans victims where both the race of the victim and offender is known is tiny. What can the limited analysis of the data that is possible tell us? We already know that the majority of these homicides are not linked to transphobia; that accounts for only four. A sample size of four is not—under any circumstances—acceptable for drawing any conclusions. Based on that, any claim that attempts to draw conclusions about the prevalence of anti-trans bias in a specific population based on the percentage of trans murders committed by people from that population should be immediately discarded. And again, to go back to Ngo’s tweet, there is no valid analytical reason to point out the race of the offenders, unless one wants to point out a problem with the people of that race. Yet, as it should be clear by now, what little data there is, in no way supports that implication. If we look at the specific cases of black trans victims, which account for 37 out of the 52 in which the race of the offender is known, we find that 34 of the offenders are also black. This is 92% of the total. A sample size of 37 is very small, and any conclusions drawn from such a sample should always be treated very carefully. Doing some algebra with the equation presented earlier, we could adjust the expected proportion of black offenders from the 0.50 consistent with maximum variation and with the largest sample size, to the 0.82 that would be consistent with the FBI data. This would lower the required sample size to 56 for a margin of error of 10%, which is still larger than the available sample of 37. Some more algebra shows that a sample size of 37 would give a margin of error of 12.4%. Finally, this means that the 92% of black perpetrators is statistically no different from the 82% for the overall population of black homicide victims. Furthermore, in the case of white trans victims, there are no black perpetrators, though in this case the sample size is of seven, so no real conclusions can be drawn from this.
As I hope, all these explanations and bits of statistical theory have made clear, the complexity of this issue. As such, any statistical analysis of this topic requires much more depth than citing one or two nominally factual data points. While both of Ngo’s claims are technically true, the proper context paints a widely different picture. As I said earlier, there are only two possibilities for what Ngo (and later his defenders) were doing. The first is that they were unaware of the complexities and really thought that Ngo’s tweet was a valid refutation of the original claim about violence against trans people. I find this hard to believe, at least in the case of Hoff Sommers and Boghossian, given that they are accomplished academics, and I believe them to be smart individuals, even if I disagree with them. This being the case, I do not know what else their actions could amount to other than feigned ignorance in bad faith. In the case of Ngo, I have no idea what his academic credentials are, if any. For that reason, he could just be ignorant of the issues. However, as it has been previously reported, Ngo has, in the past, shown to have no qualms with biased reporting. I fail to see how this could be any different. While provocateurs like Ngo will always be around, I do think it is time that public intellectuals on the right-of-center be held to much higher standards, especially given that many of them like to present themselves as dispassionate analysts and purveyors of facts.
Néstor de Buen holds an M.A. in social sciences from The University of Chicago. He has previously written at Quillette. He can be reached at email@example.com or on Twitter @nestor_d