6 problems you should know before writing about Facebook’s emotional contagion study

In a Facebook study published this week, Facebook manipulated all of their U.S. many of their users’ News Feeds by omitting 0-90% of posts containing either positive or negative content over the course of a week in 2012. They reported that those users wrote fewer positive and negative words (respectively) in their own posts, concluding that Facebook is a medium on which emotions spread, a case of “emotional contagion” using their technical term.

Here’s what you need to know:

On average, no emotion actually spread

The number of positive words in their average user’s posts decreased from 6 words to… 6 words.

The first major omission in the study is the lack of individual-level statistics. While they reported aggregate numbers such as having analyzed ”over 3 million posts” totaling “122 million words” made by their “N = 689,003″ users, and the study’s implication for “hundreds of thousands of emotion expressions,” they omitted any discussion of whether and how individuals were affected in any meaningful way.

From their numbers, the average user wrote 4.5-5 posts totaling 177 words during the experimental week. Only 3.6% of those words — so about 6 words — were “emotional,” and they found that by omitting about half of emotional posts from user’s News Feeds that percentage would go down by 0.1% or less. A 0.1% change is about 2/10ths of a word.

For most of their users, there was not even close to a measurable effect.

(The study did mention the Cohen’s d statistic of ’0.02′ which is another way to say that there was an aggregate effect but basically no individual-level effect.)

The study has no test for external validity (was it about emotions at all?)

An important part of every study is checking that what you’re measuring actually relates to the phenomenon you’re interested in. This is called external validity. The authors of the Facebook study boasted that they didn’t think of this.

The paper quixotically mentions that “no text was seen by the researchers” in order to comply with Facebook’s agreement with its users about how it will use their data.

They didn’t look at all?

That’s kind of a problem. How do you perform a study on 122 million words and not look at any of them?

Are the posts even original, expressive content? The users might be sharing posts less (sharing is sort of like retweeting) or referring less to the emotional states of friends (“John sounds sad!”). The words in a post may reflect the emotions of someone besides the poster!

To classify words as “positive” or “negative” the study consulted a pre-existing list of positive and negative words used throughout these sorts of social science research studies. This comes with some limitations: sarcasm, quotation, or even simple negation completely cut out the legs under this approach. I actually think in aggregate these problems tend to go away, but only when you have a large effect size.

The whole of Facebook’s reported effect on emotion could be due to one of the many limitations of using word lists as a proxy for emotion. They needed to demonstrate it wasn’t.

Methodological concerns

This study is not reproducible. While most research isn’t ever reproduced, that it could be provides a check against the fabrication of results (and sometimes that’s how fabricators are caught). Facebook provides the only access to a network of this size and shape. It is unlikely they would provide access to research that might discredit the study.

The study also uses a strange analysis. Their experimental design was 2 X 9-ish (control or experiment X 10-90% of posts hidden), but they plugged the two variables into their linear regression in two ways. The first became a binary (“dummy”) variable in the regression, which is right, but the second become a weight on the data points rather than a predictor. That’s an odd choice. Do the results come out differently if the percentage of posts hidden is properly included in the regression model? Did they choose the analysis that gave the results they wanted to see? (This is why I say “about half of emotional posts” above, since the analysis is over a weighted range.)

Informed consent

Finally, there’s the problem of informed consent. It is unethical to run experiments on people without it. The paper addresses legal consent, in the sense that the users agreed to various things as a pre-condition for using Facebook. Though being manipulated was probably not one of them (I don’t know what Facebook’s terms of service were in early 2012 unfortunately).

Certainly the consent didn’t reach the level of informed consent, in which participants have a cogent sense of what is at stake. There’s a great discussion of this at Slate by Katy Waldman.

Facebook’s users have a right to be outraged over this.

Keep in mind though that there are different ethical obligations for research versus developing a product. It could be ethical for Facebook to manipulate News Feeds to figure out how to increase engagement while at the same time being unethical for a research journal to publish a paper about it.

Comments are closed.