East Hunter
Dec 31, 2025
Some of the less offensive racial stereotypes are about food preferences. You heard them all: Mexicans eat tacos and burritos, the Chinese cats and dogs, the French snails and frogs. I always found it hilarious that Black people supposedly eat fried chicken, watermelons and drink cheap fruit-flavored beverages (“purple drank”) because I like these too.
But is this all true?
I decided to check this empirically [1] in a large public dataset of Americans. Participants self-reported if they were non-Hispanic White, Asian, Black or Hispanic/Latino. All four races were reasonably well represented.
In the amazing Mike Judge series Silicon Valley, some startup founders try to build a “Shazam for food”, an app that lets you can take a picture of your meals and it uses AI to automatically code it and calculate nutrients. There are apps like this now, which are OK but far from perfect. In the absence of a perfect AI, there is no very good practical way to record what people eat. Ideally, they would measure with a scale and accurately log everything they eat, but you can’t realistically recruit people conscientious enough to go through with this. Another way is to do a 24-hour dietary recall, that is, once a day ask people what they ate. Another is the Food Frequency Questionnaire (FFQ), where people report what they usually eat. It looks like this:
My dataset has this FFQ, recording whether and how frequently participants eat a large number of food items.
As you can see, the top items are typical ethnic food, probably preferred either by Asians (stir-fried items, Asian noodles, dumplings) or Hispanics (Mexican rice, tortillas, beans). Eating the skin of potatoes seems to vary a lot across racial groups, this is unexpected.
For each food-ethnic group combination, I calculated a race preference score:
In plain language, I calculate how much the preference of a certain race for a food item differs from the mean preference of the other 3 races. Dividing by the grand mean frequency and multiplying it by 100 makes it approximately a % difference. For example, the White race preference score for “Wine” of 170 means that White people consume about 170% more wine (not just 70% more!) than the other three races. Ranking the food items by this race preference score highlights the “Whitest”, “Blackest” etc. foods.
Asian and Hispanic foods are unsurprising, these ethnicities eat a lot more of the food items you’d stereotypically expect them to like:
Asians eat Asian food and Hispanics eat South-Central American foods, no surprises here. What is the Blackest food then?
Of course it’s fried chicken. Watermelon is not in the FFQ, but unusual fruit juices and non-diet beverages among the top items somewhat confirms that American Blacks like very sweet, fruity beverages. In line with the Bennett review I linked above, Blacks don’t seem to have a very healthy diet, many top items are deep-fried or fatty meals and sweets.
What about White people?
Whites stand apart from other ethnicities mainly by drinking more alcohol. All alcohol items (wine, liquor and beer) appear among the top 20. The other things that stands out are dairy items: yoghurt, cottage cheese, cheese, butter, milk etc. This is most likely because there are very big race differences in the rates of lactose intolerance: it is rare among Europeans but very common outside of Europe. The basal human phenotype is that you can’t drink milk anymore when you stop breastfeeding. In regions where people had animals they could milk (goats, sheep, cows) it was very advantageous during famines to be able to drink this milk, even when you were an adult. Therefore, multiple hotspots across the globe independently evolved lactase persistence (the ability to drink milk past infancy), but it didn’t happen everywhere. Among Americans, the rate of lactose intolerance is several times higher if they ancestors came from outside of Europe. Interestingly, Asians have practiced agriculture for a long time but they still have the highest rates of lactose intolerance.
For this reason, it is important that we go beyond single food items and look at the big picture: how well can we predict race from self-reported food preferences?
For this, I used a simple random forest classifier in R. All food preferences were used as predictors and the four-category race variable was the outcome. In a holdout sample, 82% of people were correctly identified as members of their race based on food preferences alone, over twice the random guessing rate of ~40%. Cohen’s Kappa for this is 0.74 which is a pretty good value.
Confusion matrix in the holdout sample (food preference-predicted and actual race)
Note that most of the misclassification happens between Blacks and Whites. This shows what we saw before, that Asians and Hispanics are the ones with particularly distinctive diets. This is also apparent when we plot the most important variables: these are almost all ethnic food items, showing that whether you eat something very ethnic like “tortilla as side dish” or “stir-fried vegetables” provides the model with the most information about your probable race.
To establish if ethnic food stereotypes are accurate, I created a poll which a friend with a large following posted on Twitter. In this poll, I copied the name of each food item I had and asked participants to rate if, in their view, “White people”, “Black people”, “Asian people”, or “Hispanic people” prefer this food the most. If they were unsure, they were prompted to give a random guess. I received 25 answers, which was a positive surprise given close to 200 quite weird questions and no compensation. Most participants were from the US or some other English-speaking country, with a college degree.
The first thing I noticed is that people believe food preferences are much more concentrated within races than they really are. To quantify this, I calculated a pseudo-R^2 of respondent-rated food preferences. The most extreme case would be if everybody said the same race likes a food item the most: for example, if all 25 respondents agreed that Black people like fried chicken the most. I wanted to be R^2=1: race explains all the food preference. The other extreme (R^2=0) would be a flat histogram: an exactly equal number of people saying that either Whites, Blacks, Asians or Hispanics like fried chicken the most.
The actual R^2 value was defined as:
In plain words, for each food item, I summed up the deviation of preference ratings from a flat histogram (=what % of people say each race prefers that food, and how different is this from 25%?) as the numerator and divided this with the most extreme case (everybody agrees Blacks like fried chicken the most) too see how close it is to the truth.
For the current case with K=4 races:
Survey respondents guessed a bit better than random (r=0.15, p=0.09) which food items showed the most race-specific preference patterns. As I said, they usually believed preference for food items is much more concentrated across the races than they really were.
For most food items, the concentration of preference within a specific race was actually quite low (Actual R^2). However, external respondents often – falsely – agreed with each other (Predicted R^2) that a single race prefers a food item the most.
To really evaluate stereotype accuracy, for each race I correlated self-reported food preferences (% extra preference compared to the other races, as above) with community-predicted preference. This community-predicted preference was the proportion of respondents saying a certain race prefers that food, minus the mean share of the others. This was to reproduce the logic of self-report data, which also looked at the difference between the preference of each race and the mean preference of the others. For example, if out of four respondents three said Black people like fried chicken the most and fourth said it was Asians, Black preference was 0.75-(0.25+0+0)/3=0.67, Asian preference 0.25-(0.75+0+0)/3=0, while White and Hispanic each 0-(0+0.25+0.75)/3=-0.33. Thus, the scale of self-reported and community-predicted preferences are not comparable but they follow a similar logic. If stereotypes are accurate, we would expect that Black people really said they like this food more than the other three races.
For Hispanics and Asians, stereotype accuracy is very strong:
These correlations at r~0.8 between community-predicted and actual food preferences are about as strong as you will ever see in the social sciences. People correctly guess which food items are preferred by Asian and Hispanic Americans. On the other hand, this is not that impressive, because both ethnic groups have a distinctive cuisine with well-known items. In the case of Hispanics, even the language of the food names is often different. Maybe it’s surprising to some that Asians and Hispanics really eat their ethnic food, but it’s not that big a shock.
Even for Blacks, there is strong stereotype accuracy, with r=0.53. Fried chicken is correctly predicted to be the most stereotypical Black food, but this is not the end of the story, other reasonably Black items are also seen as such by respondents.
I find the White chart the most impressive:
Unlike Asians and Hispanics, American Whites don’t have a plainly obvious ethnic cuisine. The Asian and Hispanic correlations arguably just reflected that these groups have an easily identifiable ethnic cousine, and the Black correlation was boosted by the Asian/Hispanic food items in the bottom left quartile, correctly guessed to be rarely eaten by Blacks. But here we see a nice linear relationship between predicted and actual preference, and a correlation of 0.6. It is interesting to note that the White preference for alcohol consumption is not widely known – look at how far wine, liquor and beer are above the regression line.
In summary, diet is one of the many things that shows ethnic/racial differences in America. People overestimate how much the races differ in their consumption of single food items, but they get it right which food items are typically consumed by which race, and a machine learning approach can classify the race of respondents with over 80% accuracy.
[1] You can also look up the review by Bennett et al (2022).
Source (Archive)
Dec 31, 2025
Some of the less offensive racial stereotypes are about food preferences. You heard them all: Mexicans eat tacos and burritos, the Chinese cats and dogs, the French snails and frogs. I always found it hilarious that Black people supposedly eat fried chicken, watermelons and drink cheap fruit-flavored beverages (“purple drank”) because I like these too.
But is this all true?
I decided to check this empirically [1] in a large public dataset of Americans. Participants self-reported if they were non-Hispanic White, Asian, Black or Hispanic/Latino. All four races were reasonably well represented.
In the amazing Mike Judge series Silicon Valley, some startup founders try to build a “Shazam for food”, an app that lets you can take a picture of your meals and it uses AI to automatically code it and calculate nutrients. There are apps like this now, which are OK but far from perfect. In the absence of a perfect AI, there is no very good practical way to record what people eat. Ideally, they would measure with a scale and accurately log everything they eat, but you can’t realistically recruit people conscientious enough to go through with this. Another way is to do a 24-hour dietary recall, that is, once a day ask people what they ate. Another is the Food Frequency Questionnaire (FFQ), where people report what they usually eat. It looks like this:
My dataset has this FFQ, recording whether and how frequently participants eat a large number of food items.
First question: The consumption of which food items varies the most across races?
For this, I built a regression model for each food item with consumption frequency as the dependent and race as the single predictor variable. R^2 from these models tell us which food items have the most race-dependent consumption patterns. Here is the list:As you can see, the top items are typical ethnic food, probably preferred either by Asians (stir-fried items, Asian noodles, dumplings) or Hispanics (Mexican rice, tortillas, beans). Eating the skin of potatoes seems to vary a lot across racial groups, this is unexpected.
Second question: Which food items are the most preferred by a single specific race?
You can get a high R^2 value by having two races prefer a food item and the other two almost never eat it at all. We are going one step farther now: which are the food items which are only preferred by a single specific race and not the others? In other words, which are the most White, Black, Asian and Hispanic foods?For each food-ethnic group combination, I calculated a race preference score:
In plain language, I calculate how much the preference of a certain race for a food item differs from the mean preference of the other 3 races. Dividing by the grand mean frequency and multiplying it by 100 makes it approximately a % difference. For example, the White race preference score for “Wine” of 170 means that White people consume about 170% more wine (not just 70% more!) than the other three races. Ranking the food items by this race preference score highlights the “Whitest”, “Blackest” etc. foods.
Asian and Hispanic foods are unsurprising, these ethnicities eat a lot more of the food items you’d stereotypically expect them to like:
Asians eat Asian food and Hispanics eat South-Central American foods, no surprises here. What is the Blackest food then?
Of course it’s fried chicken. Watermelon is not in the FFQ, but unusual fruit juices and non-diet beverages among the top items somewhat confirms that American Blacks like very sweet, fruity beverages. In line with the Bennett review I linked above, Blacks don’t seem to have a very healthy diet, many top items are deep-fried or fatty meals and sweets.
What about White people?
Whites stand apart from other ethnicities mainly by drinking more alcohol. All alcohol items (wine, liquor and beer) appear among the top 20. The other things that stands out are dairy items: yoghurt, cottage cheese, cheese, butter, milk etc. This is most likely because there are very big race differences in the rates of lactose intolerance: it is rare among Europeans but very common outside of Europe. The basal human phenotype is that you can’t drink milk anymore when you stop breastfeeding. In regions where people had animals they could milk (goats, sheep, cows) it was very advantageous during famines to be able to drink this milk, even when you were an adult. Therefore, multiple hotspots across the globe independently evolved lactase persistence (the ability to drink milk past infancy), but it didn’t happen everywhere. Among Americans, the rate of lactose intolerance is several times higher if they ancestors came from outside of Europe. Interestingly, Asians have practiced agriculture for a long time but they still have the highest rates of lactose intolerance.
Third question: How different overall is the diet of American racial groups?
Other than Asians eating Far Eastern dishes and Hispanics South-Central American foods, at first glance the race differences in diet don’t look that big. But of course we must be aware of the univariate fallacy: there are many human characteristics, and a small difference in each of them can balloon in a very different overall phenotype. For example, people will tell you that men are women are not that different in any single characteristic (unless it’s upper-body strength), or that there are only small race differences in the frequency of any genetic variant (unless it’s Duffy negativity). But this is what we call the univariate fallacy: objects differ from each other along multiple characteristics and if we use all of them like we should, we can classify them much better. In reality you can classify sex or race with high accuracy if you look at a larger pool of psychological or genetic characteristics. For example, having a different English accent (an immediately noticeable difference) is the sum of many small differences in how we create phonemes.For this reason, it is important that we go beyond single food items and look at the big picture: how well can we predict race from self-reported food preferences?
For this, I used a simple random forest classifier in R. All food preferences were used as predictors and the four-category race variable was the outcome. In a holdout sample, 82% of people were correctly identified as members of their race based on food preferences alone, over twice the random guessing rate of ~40%. Cohen’s Kappa for this is 0.74 which is a pretty good value.
Confusion matrix in the holdout sample (food preference-predicted and actual race)
Note that most of the misclassification happens between Blacks and Whites. This shows what we saw before, that Asians and Hispanics are the ones with particularly distinctive diets. This is also apparent when we plot the most important variables: these are almost all ethnic food items, showing that whether you eat something very ethnic like “tortilla as side dish” or “stir-fried vegetables” provides the model with the most information about your probable race.
Fourth question: How accurate are stereotypes about ethnic food preferences?
Some people may find stereotypes about ethnic food preferences offensive. I don’t understand this - what’s wrong about liking fried chicken? – but even if you do, it doesn’t matter very much because the stereotype is either correct or not, which is what matters. Stereotypes are usually quite correct – Lee Jussim famously wrote that this is the best replicated finding in social psychology – which is unsurprising because humans evolved to observe each other accurately.To establish if ethnic food stereotypes are accurate, I created a poll which a friend with a large following posted on Twitter. In this poll, I copied the name of each food item I had and asked participants to rate if, in their view, “White people”, “Black people”, “Asian people”, or “Hispanic people” prefer this food the most. If they were unsure, they were prompted to give a random guess. I received 25 answers, which was a positive surprise given close to 200 quite weird questions and no compensation. Most participants were from the US or some other English-speaking country, with a college degree.
The first thing I noticed is that people believe food preferences are much more concentrated within races than they really are. To quantify this, I calculated a pseudo-R^2 of respondent-rated food preferences. The most extreme case would be if everybody said the same race likes a food item the most: for example, if all 25 respondents agreed that Black people like fried chicken the most. I wanted to be R^2=1: race explains all the food preference. The other extreme (R^2=0) would be a flat histogram: an exactly equal number of people saying that either Whites, Blacks, Asians or Hispanics like fried chicken the most.
The actual R^2 value was defined as:
In plain words, for each food item, I summed up the deviation of preference ratings from a flat histogram (=what % of people say each race prefers that food, and how different is this from 25%?) as the numerator and divided this with the most extreme case (everybody agrees Blacks like fried chicken the most) too see how close it is to the truth.
For the current case with K=4 races:
Survey respondents guessed a bit better than random (r=0.15, p=0.09) which food items showed the most race-specific preference patterns. As I said, they usually believed preference for food items is much more concentrated across the races than they really were.
For most food items, the concentration of preference within a specific race was actually quite low (Actual R^2). However, external respondents often – falsely – agreed with each other (Predicted R^2) that a single race prefers a food item the most.
To really evaluate stereotype accuracy, for each race I correlated self-reported food preferences (% extra preference compared to the other races, as above) with community-predicted preference. This community-predicted preference was the proportion of respondents saying a certain race prefers that food, minus the mean share of the others. This was to reproduce the logic of self-report data, which also looked at the difference between the preference of each race and the mean preference of the others. For example, if out of four respondents three said Black people like fried chicken the most and fourth said it was Asians, Black preference was 0.75-(0.25+0+0)/3=0.67, Asian preference 0.25-(0.75+0+0)/3=0, while White and Hispanic each 0-(0+0.25+0.75)/3=-0.33. Thus, the scale of self-reported and community-predicted preferences are not comparable but they follow a similar logic. If stereotypes are accurate, we would expect that Black people really said they like this food more than the other three races.
For Hispanics and Asians, stereotype accuracy is very strong:
These correlations at r~0.8 between community-predicted and actual food preferences are about as strong as you will ever see in the social sciences. People correctly guess which food items are preferred by Asian and Hispanic Americans. On the other hand, this is not that impressive, because both ethnic groups have a distinctive cuisine with well-known items. In the case of Hispanics, even the language of the food names is often different. Maybe it’s surprising to some that Asians and Hispanics really eat their ethnic food, but it’s not that big a shock.
Even for Blacks, there is strong stereotype accuracy, with r=0.53. Fried chicken is correctly predicted to be the most stereotypical Black food, but this is not the end of the story, other reasonably Black items are also seen as such by respondents.
I find the White chart the most impressive:
Unlike Asians and Hispanics, American Whites don’t have a plainly obvious ethnic cuisine. The Asian and Hispanic correlations arguably just reflected that these groups have an easily identifiable ethnic cousine, and the Black correlation was boosted by the Asian/Hispanic food items in the bottom left quartile, correctly guessed to be rarely eaten by Blacks. But here we see a nice linear relationship between predicted and actual preference, and a correlation of 0.6. It is interesting to note that the White preference for alcohol consumption is not widely known – look at how far wine, liquor and beer are above the regression line.
In summary, diet is one of the many things that shows ethnic/racial differences in America. People overestimate how much the races differ in their consumption of single food items, but they get it right which food items are typically consumed by which race, and a machine learning approach can classify the race of respondents with over 80% accuracy.
[1] You can also look up the review by Bennett et al (2022).
Source (Archive)
