Nearly 600 sexual interests from the Big Kink Survey, plotted by how common they are, how taboo they are, and who likes them more. Hover a dot for details; click to pin it.
This is from the Big Kink Survey: 562,816 people aged 14–42. The raw sample is kinkier than the general population (it's people who chose to take an enormous kink survey, so of course it is), which is why everything here is weighted (“hardweight”) to match population benchmarks on age, sex, LGBTQ status, politics, ethnicity, BMI, mental health, abuse history, religion, partner count, relationship status and porn use. Weighting this hard isn't free - the weights are so unequal that those 563k people are worth about as much as 31,075 people sampled cleanly (the effective n). Every dot has at least 200 people into it.
On ordered axes (age, politics, partners…) each dot is the average of the people into it, weighted like everything else. Partner count uses the median, because a few people with hundreds of partners would drag every mean up. Averages squish - fan ages only span 24.8 to 32.9, so a couple of years is a big gap - and they hide shape, so pin an item to see the full curves.
A level needs at least 20 effective endorsers before it gets plotted. The weights are what make cells thin (563k people are worth about 31k effective ones), so when a cell comes up short, the build relaxes the weights for that one cell: cap them at 10, then 5, 3, 2, 1.5, and finally use the plain unweighted count, stopping at the first cap that reaches 20. Each step costs accuracy. Capping at 10 moves a typical rate by about 5%, capping at 2 moves it about 13%, and going fully unweighted moves it about 58% (the raw sample is kinkier), so each step only gets used when the one before it fails. Relaxed cells are drawn with a lighter fill and named in the tooltip. A cell that can't reach 20 endorsers even unweighted is drawn hollow and never used to position a dot. About 11% of cells needed relaxing; 2 in 38,000 stayed hidden.
By default the centre line on a skew axis means 1:1, equally common in both groups. That's a bad baseline for gender identity, because trans and nonbinary respondents report more interest in almost everything - the median item already sits around 2×, so nearly every dot lands on the trans side and the actually interesting question (which kinks are especially trans-skewed?) gets buried. Re-centring moves the centre line to the median item. Ticked, the middle means no more skewed than the typical item.
Groups differ in more than one way at once. If heavy porn users are more into something, is that about porn, or is it a sex difference in disguise? This option tries to pull those apart. It swaps every curve from the rate people actually reported to a model-adjusted rate: one weighted logistic regression per item (same population weights as everything else), giving the rate at each level of a variable for people who are otherwise alike on sex, age, gender identity, ethnicity and religious upbringing. Those five are adjusted for each other. The lifestyle and trait variables (orientation, the two attraction measures, politics, mental health, partner count, body weight, porn use) are each adjusted for the five, but deliberately not for each other - I didn't want the sex comparison quietly conditioned on porn habits.
It matters most for the attracted-to axes. Unadjusted, "attracted to masculine" mostly just means "is a straight woman or a gay man," so the raw ratio is basically the sex axis flipped. Adjusted, it turns into a much better question: within a sex, does wanting masc vs fem partners predict the kink? To be clear, this is a statistical adjustment, not a causal claim. It's off by default because the unadjusted rates are the ones you can check against the explorer.
Some items were asked as a 1–5 scale and some as a plain checkbox, and those aren't the same instrument - a checkbox is blunter. So where do you draw the line on a scale so that it means the same thing as a tick? I tested this directly with the Scale Explorer experiment, which randomly assigned people to get the same items as a checkbox, a 4-point scale or a 7-point scale. The share of people who tick a plain checkbox matches the share who rate an item 3 or more out of 5. So that's the cutoff here: scale items count you as into it at 3+, not at "mild interest" (2+).
Hardweight is the Big Kink Survey's aggressive population calibration: raked to census/benchmark targets on age × sex × LGBTQ, politics, ethnicity, BMI, mental health (including anxiety), abuse history, religious upbringing and current religiosity, partner count, relationship status and porn use (the explorer's methodology has the details). 562,816 respondents aged 14–42, and every dot has at least 200 endorsers.
Weighting this hard has a price. The weights are very unequal, which makes those 563k people worth an effective sample of n ≈ 31,075 for the weighted percentages (design effect ≈ 18.1). That's the number quoted above the map. It's plenty for a map, but treat any single item's weighted estimate like it came from a 31k-person survey, not a 563k one. The emotion signatures and the endorser counts in the tooltip are unweighted, so the raw n applies to those.
These come from two different questions, and they're easy to mix up. Attracted to: masc–fem is the seven-point item "you're more sexually attracted to people who appear visually … totally feminine → totally masculine". Attracted to: women–men combines the genital question ("you're sexually attracted to people with … penises / vaginas", which the explorer carries as orientation once crossed with the respondent's sex) with that presentation item: women = straight men and gay women who also lean feminine on the slider; men = straight women and gay men who lean masculine; both / mixed = bisexual respondents, plus anyone whose two answers disagree or who sits right at the midpoint. Unadjusted, both are mostly the sex axis flipped, because most people attracted to men are women. The adjusted view is where they get interesting, since there it's asking whether the partner someone wants predicts the kink within a sex.
Two things to keep in mind. First, the demographic axes are correlated with each other: trans/nonbinary, younger, non-straight, more-mentally-ill and heavier-porn-using respondents all over-report a similar cluster of items, so thirteen axes are not thirteen independent findings. (The adjusted view exists precisely to help separate them.) Second, the ethnicity axis also picks up ordinary own-group attraction - non-white respondents rate the non-white attraction checkboxes higher, which is not a kink finding. And the weighting makes some groups small in effective terms: trans and nonbinary respondents are 72,317 people but only about 1,300 effective ones, which is why thin cells get hidden rather than shown.
Tabooness is how taboo a thing is believed to be. It's a judgement about society - not the rater's own arousal, and not rated by this survey's respondents. It merges two separate surveys I ran: one put items on a 0–5 "not taboo → extremely taboo" scale (n ≈ 1,100–2,100 ratings per item), the other asked a 0–100 slider, "how taboo is X?" (n ≈ 160–2,000 per item). Items were matched across the surveys and to this map by meaning, not wording (one survey's "inserting things into the urethra" is this map's "Sounding"), and every match was hand-checked. Where both surveys cover an item they agree closely (r = 0.94 across the 57 shared items), so the 0–5 scale was linearly calibrated onto the 0–100 one and the two were combined by inverse-variance weighting. 57 items draw on both surveys, 161 on the slider alone and 38 on the 0–5 alone.
Caveat: the taboo raters are a different, much smaller and more online sample than the 563k who reported their own interests. And where a survey asked separately about male and female (or giving and receiving) versions of the same thing, those were averaged into one number.
The emotion signature (in the tooltip and detail panel) comes from two questions: which emotion you most want to feel during sex, and which you most want the other person to feel - one pick each, out of 22. For each item I show the emotion its fans over-pick most distinctively, with its lift over the base rate ("3.7× base" = fans pick it 3.7× as often as everyone).
The obvious ways of doing this don't work. Taking fans' most common emotion returns Eagerness or Love for nearly every item, because those dominate the base rates. Taking raw lift returns Despair for nearly every item, because it's so rare that any dark fetish inflates it. So winners are ranked on shrunken log-lift, centred against each emotion's typical enrichment across all 590 items, and have to clear lift ≥ 1.1, z ≥ 2 and ≥ 25 fan picks.
Signatures are computed within each sex: an item's male fans are compared to male answerers, and its female fans to female answerers. Otherwise sex differences in the base rates (women pick Powerlessness twice as often as men; men pick Love and Power more) would masquerade as fetish signal. Unweighted, and correlational. The emotion filter above the map matches an item if any of its four cells (men/women × feel/other) carries that emotion; a "—" cell means nothing cleared the bar for that group.
One more honest caveat: checkbox items were only shown to people who opened that category's page in the survey, so their denominator is "people who opened the page", not everyone. The calibration fixes what a tick means, but not who saw the box - checkbox prevalences for obscure categories run somewhat hot relative to the scales.