Digital Discrimination: The Case of Airbnb.com

August 28, 2014 10:33 AM

[...]non-black [AirBnb] hosts earn roughly 12% more for a similar apartment with similar ratings and photos relative to black hosts.

Benjamin Edelman and Michael Luca examine race discrimination in the context of AirBnB.

Despite the potential of the internet to reduce discrimination, our results suggest that social platforms such as Airbnb may have the opposite effect. Full of salient pictures and social profiles, these platforms make it easy to discriminate—as evidenced by the significant penalty faced by a black host trying to conduct business on Airbnb.

Related: Nancy DiTomaso on How Social Networks Drive Black Unemployment

posted by threeants (42 comments total)

"non-black [AirBnb] hosts earn roughly 12% more"?

Orr, like the summary of the paper says:

"we show that non-black hosts charge approximately 12% more than black hosts for the equivalent rental"?
posted by I-baLL at 10:47 AM on August 28, 2014 [9 favorites]

Why does Airbnb make the race of the host public?
posted by roomthreeseventeen at 10:48 AM on August 28, 2014 [2 favorites]

probably the host does so themselves, using the "salient pictures and social profiles"
posted by rebent at 10:52 AM on August 28, 2014 [1 favorite]

"non-black [AirBnb] hosts earn roughly 12% more"?

Orr, like the summary of the paper says:

"we show that non-black hosts charge approximately 12% more than black hosts for the equivalent rental"?

I'm not too well-versed in this, but doesn't mainstream economics believe that the primary reason someone wouldn't charge 12% more is because the market doesn't support it?
posted by threeants at 10:52 AM on August 28, 2014 [5 favorites]

The host/landlords put pictures of themselves up.
posted by mrbigmuscles at 10:53 AM on August 28, 2014

roomthreeseventeen: It doesn't. The people doing the study downloaded photos of the rooms and had people go through them trying to see if they could identify the race of the owners of the property.
posted by I-baLL at 10:54 AM on August 28, 2014 [1 favorite]

Just about every AirBnB listing I've seen has a photo of the host/owner.
posted by kmz at 11:03 AM on August 28, 2014

One thing I've wondered about is how LinkedIn's status as a shadow resume affects racial discrimination in hiring. US job applicants are generally highly discouraged from putting a photo on their resumes (in contrast to the norm in many European countries), but now through LinkedIn, hiring managers or HR folks can see the personal photo of most professional applicants. I wanted to include something about this in the OP but I was really surprised to find that almost nothing seems to have been written about it.
posted by threeants at 11:07 AM on August 28, 2014 [18 favorites]

"Indeed, the last column of Table 1 indicates that black hosts’ properties tend to be located in inferior locations, and have properties that look worse (based on listing photos)."

That would, indeed, explain a 12% price gap.
posted by corb at 11:12 AM on August 28, 2014 [2 favorites]

My understanding is that hiring entities are very happy not to have pictures on the resume, because it opens them up to potential liability in terms of discrimination charges, but I wonder if LinkedIn photos provide the ability to discriminate while at the same time affording plausible deniability?
posted by threeants at 11:12 AM on August 28, 2014

"Indeed, the last column of Table 1 indicates that black hosts’ properties tend to be located in inferior locations, and have properties that look worse (based on listing photos)."

That would, indeed, explain a 12% price gap.

The authors controlled for these factors.
posted by threeants at 11:14 AM on August 28, 2014 [11 favorites]

This was on askMeFi, and I didn't want to engage into the discussion, because it wasn't appropriate. But now it's here, I'd like to comment that this is a study of New York City. I don't believe the results apply 1:1 to the rest of the world. Not to say bigots only live in NYC, that is obviously untrue. But there are many different variations of bigotry.
Some places will certainly have similar results, others will show more racial or religious segregation and others again will have completely different results.
The AskMeFi was about AirbnB in Copenhagen, and to be completely honest, I think black Americans would be favored in that particular context. Not so much Pakistanis, but even that would depend on a lot of other factors.
I personally worry about sorting Norwegians who only come to Denmark to drink out, since I know most Norwegians are great people, and a small group do a lot of damage. People of color or of other religions are welcome at all times.
posted by mumimor at 11:15 AM on August 28, 2014

I feel like this is one of those internet axioms-- as a thread about social science research unfolds, the probability of someone pointing out an "obvious" factor that the authors controlled for rapidly approaches 1.
posted by threeants at 11:16 AM on August 28, 2014 [19 favorites]

Wait, have they accounted for the actual physical location? Or not? I see "location rating" but are they comparing places in the same neighborhoods or just places with the same "rating"?
posted by I-baLL at 11:17 AM on August 28, 2014 [1 favorite]

Wait, have they accounted for the actual physical location? Or not? I see "location rating" but are they comparing places in the same neighborhoods or just places with the same "rating"?

Yeah I wasn't clear on this either, and it bears on corb's and threeants' comments above.

"Location" is one of six categories in which you can give 0-5 stars each when you leave a review. It sounds like they've controlled for that "location" star rating. But of course people aren't necessarily rating the location against every other location in NYC, but rather against their expectations from looking at the listing.

So the question is, did they additionally correct for the location itself, using third-party data about average rents or something? I'm not sure. If not, it could certainly skew the results.

In general though I have no trouble believing that this effect exists just from pure racial bias, since it's appeared in so many other contexts. (What was the one where they changed names on resumes to be more "black sounding" and measured response rates?)
posted by neat graffitist at 11:26 AM on August 28, 2014 [3 favorites]

I think it's especially bad when you're looking at NYC, which has places where three blocks can mean an entire neighborhood shift - particularly when AirBnB doesn't give the exact address. So people are having to make judgement calls about which neighborhoods places are actually in.
posted by corb at 11:28 AM on August 28, 2014 [2 favorites]

"We estimate the gap in rents received by
non-black and black hosts, and we show that this gap persists even when controlling for
factors such as location, reviews, and photos."
posted by drlith at 11:29 AM on August 28, 2014

threeants: "doesn't mainstream economics believe that the primary reason someone wouldn't charge 12% more is because the market doesn't support it?"

Well, mainstream economics assumes that actors in an economy behave rationally. I am not sure mainstream economics is the best place to look for an explanation of this particular phenomenon.
posted by tybeet at 11:30 AM on August 28, 2014 [3 favorites]

Yeah, no, their methods are definitely open to critique! I just think it's a little silly when people are like "but did they think of THIS?! gotcha!"
posted by threeants at 11:30 AM on August 28, 2014 [1 favorite]

I'm not sure micro-neighborhoods are super relevant in the context of AirBnB. Is someone from out of town really going to have the knowledge to identify the cupcake-shop block versus the drug dealer block within the same named neighborhood? Seems doubtful.
posted by threeants at 11:32 AM on August 28, 2014

From the "Discrimination as a market design problem" section:

we see no real need for Airbnb to highlight the host’s picture.

I wonder if there is any research on people's comfort levels with real life interactions with anonymous online people vs online people with plausibly real identities. Airbnb requires a lot of trust between both parties because there are lots of things that could go wrong (not enough towels, renters trash the place, surprise torture dungeon, etc) and not a lot of built-in ways to have them addressed. Intuitively it seems like anything that helps make the other party seem like a real person would help reduce that suspicion, but has that been tested? Seems like it's relevant to this.
posted by ghharr at 11:34 AM on August 28, 2014 [1 favorite]

Well, mainstream economics assumes that actors in an economy behave rationally. I am not sure mainstream economics is the best place to look for an explanation of this particular phenomenon.

Fair enough; I'm certainly not a big fan of many of the assumptions behind economics. But what's the alternate explanation in this case? That black people are 12% more generous? That non-black people have 12% more business savvy? It certainly seems most plausible to me that, after controlling for other factors, non-black people charge more because that's what they can get.
posted by threeants at 11:41 AM on August 28, 2014 [2 favorites]

It certainly seems most plausible to me that, after controlling for other factors, non-black people charge more because that's what they can get.

That's a bit circular isn't it? The question is "Why can non-black people charge more?" and your answer is "Because they can charge more." Yes, we know that, the question is why would they be able to get more, if all other factors are the same, or controlled for?
posted by mrbigmuscles at 11:48 AM on August 28, 2014

That's a bit circular isn't it? The question is "Why can non-black people charge more?" and your answer is "Because they can charge more." Yes, we know that, the question is why would they be able to get more, if all other factors are the same, or controlled for?

Hmm, that's a fair point. Do you have an alternate explanation other than that black people have less selling power because of discrimination? Not a snarky question, I'm genuinely interested; it's just hard for me to feel uncomfortable with the discrimination hypothesis in the absence of other credible theories.
posted by threeants at 11:52 AM on August 28, 2014

Yeah, no, their methods are definitely open to critique! I just think it's a little silly when people are like "but did they think of THIS?! gotcha!"

You may not like the tone of it, but sadly it's often exactly the right reaction. Too many social scientists do not really understand statistics -- and I say that as a former poli sci major. More recently I took a Stats 101 course at an Ivy League school, and our TAs were social science PhD students who conducted a test review session where they'd make up a "demonstration" problem, put it up on the blackboard and then literally be unable to solve it themselves.

Especially in a post-Freakonomics world, I suspect that studies like this are often done by starting with a conclusion in mind, adding controls for the first few things you think of, and then stopping while the result is still significant and handwaving the next few controls you should have applied.

In this case I share their conclusion about probable racial bias, sounds like you do too, but that should make us more suspicious of the methodology, not less.

I'm not sure micro-neighborhoods are super relevant in the context of AirBnB. Is someone from out of town really going to have the knowledge to identify the cupcake-shop block versus the drug dealer block within the same named neighborhood? Seems doubtful.

Yeah that's not really the point I was making. Let's just take Manhattan vs. the outer boroughs. Suppose that 50% of white hosts are in Manhattan but only 40% of non-white hosts are. With no controls you'd expect the the white hosts to have higher rents just from that. If you control for "location" by that star rating, you may be partly eliminating that effect but you're almost certainly not capturing all of it -- because the star rating is not "Manhattan vs Queens" so much as "actual location in Queens vs what I imagined from the profile."
posted by neat graffitist at 11:52 AM on August 28, 2014 [12 favorites]

Do you have an alternate explanation other than that black people have less selling power because of discrimination? Not a snarky question, I'm genuinely interested; it's just hard for me to feel uncomfortable with the discrimination hypothesis in the absence of other credible theories.

Looking at that article about black versus white social networks, I wonder how much of it also has to do with not really knowing the most optimal wording to signal optimal location, whether or not it is real. Like for example - how a lot of shitty neighborhoods in NYC are being "renamed" for real estate purposes. As they note there, "Greenwood Heights" might fetch a premium, where "Sunset Park" might not, even if the apartments are next door to each other.
posted by corb at 11:54 AM on August 28, 2014 [1 favorite]

Do you have an alternate explanation other than that black people have less selling power because of discrimination?

Huh? No, that's my angle too. I guess I misunderstood your comment.
posted by mrbigmuscles at 11:55 AM on August 28, 2014 [1 favorite]

Looking at that article about black versus white social networks, I wonder how much of it also has to do with not really knowing the most optimal wording to signal optimal location, whether or not it is real. Like for example - how a lot of shitty neighborhoods in NYC are being "renamed" for real estate purposes. As they note there, "Greenwood Heights" might fetch a premium, where "Sunset Park" might not, even if the apartments are next door to each other.

Yeah, IIRC there is actually a lot of literature, particularly coming out of the Global South, on geographical stigma in class mobility. I.e. having a favela address can be a dealbreaker for getting a job.
posted by threeants at 11:58 AM on August 28, 2014

Let's just take Manhattan vs. the outer boroughs. Suppose that 50% of white hosts are in Manhattan but only 40% of non-white hosts are. With no controls you'd expect the the white hosts to have higher rents just from that. If you control for "location" by that star rating, you may be partly eliminating that effect but you're almost certainly not capturing all of it -- because the star rating is not "Manhattan vs Queens" so much as "actual location in Queens vs what I imagined from the profile."

Sorry to be obtuse, but I'm trying to understand this and having a little trouble. Do you mean that the location rating left by users tends to incorporate expectations and fulfillment thereof, and thus it's much more relative than absolute and maybe not useful data to work with?
posted by threeants at 12:01 PM on August 28, 2014 [1 favorite]

And sorry I was snarky about control factors up-thread. I see that a lot in research-based threads and it often strikes me as maddening, but the discussion of methods here is really interesting and I'm glad my comments didn't manage to shut that conversation down!
posted by threeants at 12:11 PM on August 28, 2014 [1 favorite]

Fake photos and names (hosts protecting their security) are common so this is not a good guide

Why not? Presumably it's the perceived race of the host that matters.
posted by leopard at 12:40 PM on August 28, 2014 [2 favorites]

Do you mean that the location rating left by users tends to incorporate expectations and fulfillment thereof, and thus it's much more relative than absolute and maybe not useful data to work with?

Yeah, more or less. I mean, I'm sure it's a mix of things. Some tourists might never give a Queens location five stars, because in their mind a "five star" location in New York means walking distance from their Broadway show. But I'm guessing that most, like me, would give five stars if it met/exceeded their expectations for being close to the train, having restaurants nearby, whatever.

I'm not saying they shouldn't control for that rating -- I think it's good they did. I'm saying they should additionally control for absolute location relative to all of NYC, even if they just have an approximate location and not a street address. And there's certainly no shortage of freely available data that would help them do that. But it seems like they didn't. (Again, unless I missed something, I did find it a little unclear.)
posted by neat graffitist at 1:25 PM on August 28, 2014 [1 favorite]

> adding controls for the first few things you think of, and then stopping while the result is still significant

Alternatively just increase N until any difference, no matter how microscopic, becomes significant. But that's hard to do unless the subjects are flatworms.
posted by jfuller at 1:39 PM on August 28, 2014

Or freshmen answering questionnaires.
posted by jfuller at 1:41 PM on August 28, 2014 [4 favorites]

Ha looks like they forgot to control for race! Bozos!
posted by mulligan at 6:57 PM on August 28, 2014 [2 favorites]

re: LinkedIn you can always just not put your picture. I don't. In my case it's pretty obvious I'm not white from my name anyway but I know plenty of people who don't put their picture and it hasn't hurt them. (In fact, I know plenty of people who don't use LinkedIn at all and it hasn't harmed their professional prospects...)
posted by pravit at 8:28 PM on August 28, 2014 [1 favorite]

I'm not saying they shouldn't control for that rating -- I think it's good they did. I'm saying they should additionally control for absolute location relative to all of NYC, even if they just have an approximate location and not a street address. And there's certainly no shortage of freely available data that would help them do that. But it seems like they didn't. (Again, unless I missed something, I did find it a little unclear.)

I don't know that it's as much a rating relative to expectations; there's a separate accuracy rating for that. The real problem with the location rating is that the ratings are given only after the fact, once one has stayed. There is a bias introduced here due to this censorship. As an example, when I go to San Francisco, I usually stay in the cheap hotel district between Union Square and the Tenderloin. I would rate an AirBnB location there highly because it's close to transit, there will be a convenience store within a block, there's some great (but divey) Asian restaurants in the area, and so on. I might also rate an AirBnB location in North Beach equally highly; it's a really pleasant neighbourhood with great Italian restaurants, it's a reasonable walk to lots of sights, etc. Let's say I would give them both 4 out of 5.

Someone else who is less jaded about city life might like North Beach as much as I but would really not like the area near the Tenderloin; there's lots of homeless, a reputation for crime and so on. They might give the Tenderloin 1 out of 5, and North Beach 4 out of five. However, they have done ten minutes of reading about San Francisco (or even reading other AirBnB reviews) and they know they wouldn't like the Tenderloin, so they stay in North Beach.

Let's suppose that both of us go to San Francisco twice. I stay once in each place, and give them a rating of 4 for location. Someone Else stays in North Beach both times, and gives it a rating of 4 for location. Presto, both places have a location rating of 4. However, the true value of the location is based in not what people who stay somewhere think of it, but in what everybody considering staying there would think of it. If we looked at the ratings of everybody who considered each location, then North Beach would get a 4 and the Tenderloin would get a 2.5.

There are lots of things they could have done to properly control for the location - look at hotel costs, for example. Using the biased location rating doesn't cut it.
posted by Homeboy Trouble at 8:32 PM on August 28, 2014 [3 favorites]

I feel like this is one of those internet axioms-- as a thread about social science research unfolds, the probability of someone pointing out an "obvious" factor that the authors controlled for rapidly approaches 1.

This times a million. And on a related note, can we have another-- that the probability of someone pointing out correlation isn't causation also approaches one.
posted by Ned G at 1:53 AM on August 29, 2014

And on Metafilter at least, the probability that someone mocks people who say "correlation is not causation" also approaches 1.

Yet virtually every piece of highly publicized social science research that I've ever seen has been heavily flawed, with the hype more or less undermined by banalities like "statistical significance is not practical significance" and "the highlighted effect has been cherry-picked." Off the top of my head, recent FPPs about Facebook emotional manipulation, him-icanes and her-icanes, and dads who do dishes have more ambitious daughters all highlighted widely hyped studies with obvious flaws. Of course each of the posts were followed by some comments mocking the idea that someone with a PhD could possibly publish an overhyped piece of research.
posted by leopard at 7:31 AM on August 29, 2014 [1 favorite]

I wish there was a snappy way to say, "it's really hard to tease race apart from economic status."
posted by smackfu at 7:39 AM on August 29, 2014 [2 favorites]

Yeah, don't get me wrong leopard, I am large - I contain multitudes of minor annoyances at bad scientific reporting, and the correlation v causation thing is sometimes worth pointing out. But, it is a pet hate of mine to see it lazily trotted out in almost every comment section on a social science article because social sciences, and in fact (nearly?) all sciences work by inductive, rather than deductive reasoning.
posted by Ned G at 7:53 AM on August 29, 2014 [1 favorite]

I see the stats failure with bio people a lot, too. A bio paper or a social science paper is going to be 10%-75% statistics and thinking about statistics and methodology, but people manage to get out of undergrad from the best schools not knowing what a Gaussian is, somehow. I mean, in order for you to have a Kahnemann you also have to have a Tversky, and maybe a Diaconis. Radical underestimates of how much math and stats it really takes to do anything not simple. Maybe the class ratios should be flipped: 80% statistics, get some psych theory.
posted by curuinor at 1:24 AM on August 30, 2014 [1 favorite]