The Facebook Emotions Study Controversy
The emotion-manipulation study conducted by Facebook has drawn a lot of attention in the last week, much of it negative. Many DA readers are on Facebook (like so much of the online world), and many DA readers have children with Facebook accounts. Everyone is free to make up her own mind about how much it matters, but I’m a big believer in people making informed decisions, so I thought it might be helpful to have a more comprehensive post and discussion.
I’ll start by summarizing the main points. On June 2, 2014, two academics and a Facebook data scientist published a paper in the Proceedings of the National Academy of Sciences, which is the flagship journal of the most prestigious scholarly society in the United States. In this journal article, they reported the results of a study which attempted to manipulate the emotional content of Facebook News Feeds. More than 600,000 users were included in the study, with some being in treatment groups (their feeds were manipulated either positively or negatively, according to the study protocol) and some in control groups (their feeds were not).
The two requirements to be included were: (1) posting in the English language; and (2) having posted in the previous week. The study took place over seven days in January 2012.
US government-funded research that involves human subjects is required to conform to the Common Rule, which protects subjects from mistreatment and stipulates that they give informed consent before they participate. Informed consent is a technical requirement that places certain burdens of explanation and disclosure on the researcher. Not all studies can provide complete information to subjects before the experiment takes place because knowledge may alter behavior in ways that invalidate the study. In these cases, sometimes termed deceptive research, investigators are permitted to mislead subjects about the true aims of the study while they are preparing them for participation and in carrying out the data collection, but the subject must be debriefed after participation, ideally directly following participation but in all cases before the data are analyzed. Consequent to this debriefing, the subject has the right to have her data removed from the study.
Studies that do not receive federal funding are not required to conform to the Common Rule, but many private and public institutions (and some companies) voluntarily choose to meet these requirements. Versions of the Common Rule are in force in many other countries but by no means all.
The data collection and analysis for this emotion study were conducted by the Facebook data scientist. The academic authors, according to the notes provided by PNAS, limited their participation to designing the research and writing up the paper (together with the Facebook data scientist). This means that the academics were effectively siloed from the human subject part of the research, i.e., they were separated from the sources of the data and the subsequent collection and analysis. This is important because had the academics interacted directly with the subjects or their data, they would have had to submit their proposed research and their methods of obtaining informed consent to their institutions’ IRBs (Internal Review Boards). Not all academic research involving human subjects requires informed consent (there are various conditions under which consent rules can be waived), but that is the IRB’s decision to make, not the researcher’s.
Facebook, whose data scientists gathered and analyzed the data, asserts that they received informed consent from their human subjects. Their justification is that all Facebook users agree to Facebook’s terms of service conditions when they set up their accounts. These conditions currently include assent to research for internal operations purposes, but they did not include such provisions in January 2012, when the experiment was conducted, and there is no equivalent to an IRB within the company even though they conduct considerable social science research, with and without academics’ collaboration.
The bottom line: three Ph.D. social scientists conducted experimental research on Facebook users. Facebook did not obtain informed consent from the users before the research was conducted. The academic researchers designed the study (together with the Facebook data scientist) but then absented themselves until it was time to write up the paper. They submitted the project to their own IRB at Cornell University, and the IRB deemed informed consent to be unnecessary because they were not involved in data collection or analysis. Cornell University released a statement asserting:
Because the research was conducted independently by Facebook and Professor Hancock had access only to results – and not to any data at any time – Cornell University’s Institutional Review Board concluded that he was not directly engaged in human research and that no review by the Cornell Human Research Protection Program was required.
This is not the first time Facebook and academics have joined forces to manipulate user behavior for non-commercial (or not solely commercial) purposes. In 2010 they conducted a political experiment to increase voter turnout and claimed that the manipulation resulted in up to 340,000 more votes cast. That research was approved by the academic co-author’s IRB on the grounds that “minimal risk” was involved and therefore informed consent could be waived. Minimal risk is another technical category.
If this study is another “minimal risk” study, why are so many people, including many psychologists and social scientists, so outraged at what the three researchers have published? I can’t answer for them, but here are the reasons I find the research so problematic.
- We can’t know if the risk was minimal or not without knowing more about the population. We know that the median substantive effects were small, but we don’t know what the distribution of effects was, or what the outliers look like. Say 1% of the distribution (of emotion effect) lies at the upper end, with more than minuscule effects. That’s more than 6000 people. In addition, the total population of Facebook is not necessarily a “normally distributed” population, emotionally speaking. Its users definitely comprise a skewed sample of the world (and world’s English-speaking) population, and we don’t know if and/or how it’s skewed on this particular trait.
- There were no pretests or pilot studies reported. If there was an effort made to see what the overall effect on more than 600,000 people might be before carrying out the main experiment, it wasn’t reported. Failing to pretest is analogous to ignoring the hair-coloring manufacturer’s advice to pretest on a patch of skin before dumping it all on. Most people don’t pretest, and most people don’t have a reaction. But for the ones that do, the usefulness of the test cannot be overstated. Where the analogy breaks down, of course, is that when I color my hair, I’m responsible for following or not following the advice; the study subjects had their choice made for them.
- The results are not about emotion directly; they are about changes in behavior which the investigators infer is based on changes in the emotions of interest. There are any number of unrelated-to-the-study reasons (omitted variable bias for you stat types) why users might have posted more or less. To assume that changes in posting behavior reflect changes in users’ emotional states is a leap that cannot be credibly sustained without more data. Therefore, we don’t actually know how much or how little the experimental manipulation affected emotions.
- Facebook and other online companies run experiments all the time, but most of these experiments are not about manipulating unmediated, general states of mind. A study like the 2010 get-out-the-vote study is about politics. Other studies are about advertising. In these cases, users enter the study with prior expectations about being manipulated by politicians and companies, and they subconsciously adjust for that manipulation when they engage with the material. But we don’t get on Facebook thinking “hey, sometimes Facebook just fucks with my feelings for science,” so we don’t protect ourselves against that. We may well be more vulnerable in a general manipulation than in a specific one.
- Academics do research on emotions all the time. It’s an important topic for us to understand. Frequently, emotions research is deceptive research, so we can’t tell subjects exactly what we’re doing before they participate in the experiment. But we make sure they know they’re being experimented upon, and we debrief them honestly afterward. There are protocols that cover just about every kind of research social scientists conduct. Those protocols can be used for 6 people or 600,000 people. Use them.
- If Facebook is going to fuck with our feelings for science, especially if they’re doing it using the imprimatur of academic legitimacy, they need to follow academic rules, especially if they’re going to reap the benefits of that legitimacy (and offering academic research and publishing opportunities has been part of their pitch to social scientists). Academics work with IRBs. IRBs require informed consent. Get informed consent or don’t do the research. Even now, Facebook’s TOS specifies that their research is for “internal operations.” There is nothing internal about an open-access article in a prestigious journal.
- This study’s sample selection failed to exclude minors. No one should be able to experiment on minors in any way without the explicit, informed consent of their parents, guardians, or other legal agents.
I’ll close by reiterating what I said on Twitter and at my personal blog when I first heard about this study, before we had the whole story, but which still sums up my feelings as an academic and as someone who spends a lot of time and emotion online:
The history of human subject research is full of horrible examples of abuse and exploitation. As an empirical social scientist, I know how lucky I am that people are willing to spend their time, effort, and emotions, and even risk their reputations (however hard we try to anonymize their identities) to increase human knowledge. Their generosity is a gift that should be honored and respected.
Facebook just told us how badly they fucked with that gift.
I’m happy to answer specific questions about the study in the comments, so feel free to ask. And I apologize for my language. I guess I’m still angry.
Thank you for this. I’m still furious. And one additional point: I hold PNAS responsible, too, for accepting that crappy excuse for “Informed consent.”
@Shae Connot: I’m so furious I spelled my own name wrong!
Sunita, thanks for the clarification of this insane situation, especially from the perspective of academic review – it seems that the study was set up to take advantage of Cornell’s prestige while the methodology cleared them of improper practices.
I work in advertising, where data is collected and manipulated all the time, so I am somewhat cynical in regard to this situation. My immediate thought after hearing it was that at some point in the not-too-distant future, I will have my FB rep calling me to tell me that they can tailor users’ timelines in advance of a campaign to make them more emotionally receptive to ads served; e.g. if users are made to be depressed, ads for Prozac or sunny vacations would theoretically be more favourably received.
In some ways, that’s not a huge departure from the approach of behavioural advertising, where [theoretically user-allowed] cookies are mined to target a viewer’s demonstrated interest in a subject, advertiser, or topic in order to serve them more relevant ads and thus improve clickthroughs, or from contextual placements that rely on the users’ emotional investment in the surrounding topic to influence their opinions of the ads in the content environment. However, it’s the sinister and non-consensual aspect of manipulating users beforehand that would prompt me to keep my clients away. Essentially, we work on the basis of emotions driving behaviour (Hallmark card ads, anyone?), so it isn’t possible to separate the two, but the potential fallout from capitalizing on overt manipulation is prohibitively negative.
I think the outrage for me is not about what Facebook did. Facebook and its leadership have no moral compass about pretty much anything, and few are surprised anymore about how little they care about professional ethics. But I find myself less forgiving of the social scientist involved in writing up the results, the institution involved, and the journal where these results were published. There is NO excuse for this, especially given how easily those struggling with depression can be “tipped over” into harming themselves or how hyper-emotional many teens (who weren’t excluded from the study) are. In no way does this qualify as “minimal risk” in my book.
I hadn’t thought about minors. As if teenage emotion swings aren’t hard enough to cope with on their own.
Ugh, ugh, ugh.
Thank you for explaining everything so clearly. Don’t apologize for your language or anger; they highlight just how reprehensible FB and the others involved have been.
@Shae Connot: Yes, I think that’s a big failure here. The administrative procedures at PNAS are a bit unusual, in that there is no single editor of the journal (there is a board overseeing it) and the editor for each article has a great deal of latitude. This editor is the same one who edited and approved that highly criticized article on hurricanes with female names.
@cayenne: This does have a lot of overlap with advertising techniques, obviously. And equally obviously, advertising is designed to manipulate our emotions. It’s been hard for me (and others to articulate exactly what makes this so different, but I think the generalized manipulation (i.e., not connected to a particular behavior) is part of it. And what you point to is also different: it’s one thing to have an ad campaign that appeals to your emotions. But to manipulate emotions preemptively and then advertise based on that changed emotional state feels like a step too far.
@Amber M: Yes, I agree. I expect Facebook to adhere to minimal business ethics, which I’m still not convinced they’ve done here, but my far greater discomfort is about the academics who colluded and provided legitimacy for this process. We need to guard ourselves against companies (caveat emptor and all that), but we’re not supposed to have to guard ourselves against scholars.
@Ros: Yes, the inclusion of minors is really unforgivable. Whether they were adversely affected or not, FB’s minimum age is 13, which is way below any reasonable age of consent for psychological manipulation, in my opinion.
@Darlynne: Thanks! I generally try to keep my DA posts more balanced, but I can’t stop being outraged at this. The more I think it through, the more uncomfortable I get.
I’m with everyone in disappointment/anger on this one. I ran multiple studies in grad school (but not at Cornell) where we showed various edits of film/video to gauge emotional response and we had to be extremely careful with informed consent and debriefing – all though the process I got the message over and over that this was extremely important, and that was just a repetition to what was taught in classes on the subject.
I’m still agog that Cornell is saying that their researcher is ok because he “had access only to results” when he *knew* the methodology going in – just, argh, SO completely contrary to the way I was taught ethical human subjects research should work. And ESPECIALLY with minors involved – that’s supposed to be given even more thought by researchers. It really seems a case of being willingly blind and ignoring all the ethics profs are supposed to all follow. The whole thing seems completely unethical. I can’t believe their IRB gave this a pass.
Short version: I’d expect this (somewhat) from Facebook, I did not expect this from Cornell. It’s been over a decade since I was in grad school and I STILL remember all this stuff even though it’s nothing I work with now – it’s THAT important in research.
@K.L.: Totally agree. I detoured into the research track while getting my social work Master’s degree, and assisted with lots of social studies/research. Yes, it’s social work studies, but informed consent and ethics were a HUGE part of the process when working with participants; the importance of consent and ethical research design/implementation was hammered into us all daily.
I am totally appalled that this passed IRB; I can’t even see how it could, considering that IRB was instituted to protect people from exactly these kinds of totally unethical experiments.
The fact that this is termed “minimal risk”, as well. UGH. The problem with involving people who do not know they are in a study is that you have no way of knowing if there is mental health present, suicidality, self-injurious behavior/ideation… All of which (and more) can intersect and come to a head when presented with something, yes, as “simple” as a manipulated FB timeline. This is why consent and debriefing exist.
Ugh ugh ugh. The whole thing makes me sick to my stomach. Guess what — I was going through an absolutely brutal depressive episode in Jan 2012. Will I ever know if that was affected by this? Nope. Would I have given my consent to be part of such a study? Absolutely not.
I use Facebook to see what my friends are up to and keep in touch with folks around the world. It feels like it can’t even do that correctly — it keeps defaulting back to “top stories” no matter how many times I set it to “most recent”. People will post things that I never see, then are asked if they want to pay to make their posts more visible. If I try to adjust things it will say things like “do you want to see less of this? *post by my friend*” and it’s like what, no, I just don’t want ads for games in the middle of my feed, why would you take away my friend’s posts because of that…?
Maybe I’m just not a smart bear, but it’s a tool I use and it sure doesn’t work well, but it’s also what my friends and family use. Now I’m left wondering if people in my family, my friends, my older nieces and nephews (who would have been just old enough to have Facebook accounts) were affected, or are currently being affected because who knows what they’re doing now since they’ve decided we’ve all consented to whatever they want.
I am a clinical research coordinator working at an academic institution (not the one involved in this experiment).
Based on the statement from Cornell University: ” …that no review by the Cornell Human Research Protection Program was required,” this study got “Exemption from IRB review” meaning IRB review and oversight was not required.
(More information, including the criteria for IRB review exemption, can be found on the cornell IRB website: http://www.irb.cornell.edu/documents/IRB%20Policy%202.pdf)
My guess is the study got the Exempt status from Cornell IRB based on criteria #2(iii) Observation of public behavior, since the professor did not receive any data.
Depending on how much the IRB knows at that time, this might be a correct decision from their point of view.
Since the IRB can only make their decision based on information (e.g. protocol, application form, etc) submitted by the researcher, I wonder how much information was included in that initial application that Dr. Hancock and his research team submitted to the IRB.
Did the IRB know the WHOLE design of the experiment, or just that part the Dr. Hancock will be receiving results from fb and preparing the manuscript?
Did the IRB know the data include those from minor?
I’m not defending the IRB’s decision. What fb did was clearly unethical in my opinion, and I’m angry about this whole thing as well. However, it is sad to see that this slipped through the IRB at an academic institution as well, which sole purpose of existence is to protect human (research) subjects. Sometime, unfortunately, things are not as black and white as we wish, and there are always company/people will take advantage the trust and kindness in human nature.
@K.L.: @Diana: @Connie: Thanks to you all for weighing in. I agree with Connie that we shouldn’t assume the IRB fell down on the job, they can only work with the info they have. I think it’s the overall sense of disappointment and frustration that the scholarly side of this was so messed up. I think it’s instructive that the PNAS has issued an author-affiliation revision emphasizing that UCSF is the current affiliation of one author, not the one at the time the work was conducted. I can’t remember seeing that kind of note very often. And the APA’s decision to reiterate its informed consent guidelines is unusual too.
@Lindsay: So many people have this problem, i.e., that Facebook isn’t necessarily something they enjoy but they don’t know how to disengage without losing important connections to family and friends. It’s a problem of coordination, where the more people use something the more necessary it becomes. Sure, you could set up an alternative site to bring people together (and that’s become much easier for non-technical people to do), but that entails a lot of work and a lot of organization and persuasion.
Sunita – Thank you for explaining this so well.
It’s creepy that in discovering that they can manipulate emotions, they then look at how to apply this power to voting, slanting opinions, etc. I like FB but I don’t like this whole deal.
Sunita, I think this is the single best summary of the situation and the relevant issues I’ve read. Thank you so much for this – it’s a fantastic general reference post, as well as a great discussion generator.
@Tina Gabrielle: @Robin/Janet: I’m glad it was clear, and thanks!
@trudy: I haven’t thought as much about the political experiment Facebook did in 2010 as I should have, but one of the aspects of it that we should keep in mind is that to the extent Facebook users are skewed in a specific party direction, that skew can be reflected in the results. So, for example, if Facebook users are disproportionately likely to vote Democrat over Republican (and the US user base has characteristics that suggest that), and users are equally affected by the manipulation regardless of party orientation, then the result (voter turnout in this case) is going to lead to disproportionately more votes cast for Democratic candidates.
On the emotions study, there is a 2013 Pew survey of US internet behavior that asks about Facebook usage. It shows that women and young people use Facebook disproportionately; that should have been taken into account when constructing the study sample, but there is no evidence that it was.
@Sunita
“On the emotions study, there is a 2013 Pew survey of US internet behavior that asks about Facebook usage. It shows that women and young people use Facebook disproportionately; that should have been taken into account when constructing the study sample, but there is no evidence that it was.”
I believe the Pew study was in field after the controversial FB study and results only released early this year, so they wouldn’t have been available to Prof. Hancock et. al. But I expect they could have constructed the sample similarly, or at least they could have constructed it proportionate to FB usage skews (rather than to overall membership).
@cayenne: Yes, I realize the Pew survey was conducted the following year, but the authors didn’t need the Pew data. Facebook knows who its users are, and there are multiple ways of oversampling to adjust for skewed draws (or in this case, a population that is skewed). I brought up the Pew data because it shows those of us who don’t have access to Facebook’s data how the sample was likely to skew. What is disturbing to me is that the population skew (female, young) overlaps with the group that is likely to have emotion-related issues. The researchers either didn’t think about it or didn’t care.
[ETA: I’m not talking about the effect of the skew on the data. With a sample of over 600k, the differences in gender, age, etc. may well wash out, given the size of the differences reported in the Pew study. I’m talking about the effect on the participants. Given that the population is skewed in this way, it would be easy to underestimate the predicted effect.]