Wow! Andrew Guess and coauthors have published in Science a paper very similar to the one I described six months ago, collaborating with Facebook researchers to turn off ranked news feed in favor of chronological.

View Comments

Back in the day when I used to work at a particular social media company I saw an internal report describing how ranking, or the process of changing the order of content in user feeds, was provably a good thing. This was in the face of both public pressure from scholars and commentators describing how content ranking algorithms created bias and opportunities for controversial, divisive content to go viral and other kinds of bad things.

The authors of the report had run a randomized experiment turning off ranking for a small portion of users for a few weeks. They showed that these users spent less time using the social media platform, scrolled past more stories, engaged less with meaningful content, and a whole host of other “negative” effects. Putting aside the obvious conflict of interest of some of these measurements (as in, maybe less time spent is actually a good thing, societally speaking), this is a great example of the research design not taking into account equilibrium effects.

For years the existence of the ranking algorithm has influenced both the types of content produced both by content creators and by the social media company itself. Turning off the ranking algorithm for a small subset of users for just a couple weeks is not a good hint at what the world would look like in the absence of a ranking algorithm.

If you have a ranking algorithm, there’s no penalty for just firehosing tons of content into the ether, relying on the algorithm to pick out the cream of the crop to bring to the top. If you turn off the algorithm, everyone will hate you because they are suddenly inundated with terrible content. But in a world that never had an algorithm in the first place, it’s unlikely that so much mediocre content would exist in the first place.

View Comments

1. RCTs, Causal Inference, and Internal and External Validity

(you can skip this section if you know any of the above words)

RCT (randomized controlled trials) are considered the gold standard in social science for determining causal effects - that is whether some intervention or treatment caused a change or outcome Y. The idea is pretty basic. You randomly split your participants into two groups, give one of them treatment, measure outcomes and argue that the average difference¹ between the groups represents a causal effect, since there’s absolutely no reason to expect that the groups are different from each other in any meaningful way (because they were assigned randomly).

Tech companies have taken this idea and run with it. Facebook is famous for running thousands of experiments on its users simultaneously all the time. Maybe 1% of users are seeing more posts from friends in their news feed compared to everyone else. After two weeks, the team running the experiment will look at some metrics like time spent scrolling News Feed, or number of likes or comments written. If people who see more posts from friends spend more time on Facebook, they’ll roll out the change to everyone. This is roughly how Facebook, and many other companies, “optimize” their product: someone dreams up a change, engineers build it, it gets tested for a bit, and then if the metrics look good it gets shipped.

Curiously, in the social sciences many academics would say that a “single experiment is not enough to determine policy,” which is precisely what tech companies use RCTs for (and increasingly NGOs and governments when evaluating policy programs!²). The concerns from academics usually boil down to something called “external vs internal validity.” Internal validity refers to the idea that the effect of the treatment is accurately estimated for the people in your experiment. Most RCTs have near-perfect internal validity. External validity, on the other hand, refers to the idea that the results of the experiment generalize to a larger population.

There are a bunch of reasons why external validity might be dubious in an RCT. For example, if you only did your experiment on young women, one might be suspicious that the effects generalize to older men. In this post, I want to focus on a specific kind of external validity violation, which is actually a risk for many experiments.

2. Equilibrium Effects

a risk for interpreting the results of any experiment to inform policy

The Heisenberg effect (which roughly says that one cannot know both the position and momentum of a particle at a given time) is popularly confused with the observer effect, which says that the act of measuring [the position or momentum of a particle] will actually mess with the position or momentum of that particle, since there’s no way to measure something without actually interacting with it.

In a similar way, it’s often impossible to know the equilibrium effect of rolling out a treatment to a large audience based on a randomized experiment involving a smaller sub-audience. This is because for most people in the actual world, the status quo hasn’t changed, even if it’s different for people participating in the experiment.

Call your treatment Z. By construction, everyone in your experiment lives in a world where not-Z is the norm. Your experiment randomly gives Z for half the actors in your experiment. But importantly, you can only measure the effect of “having Z when most people have not-Z.” This is a different effect of “having Z when everyone else also has Z,” which would be the case if you were to “roll out” the policy globally.³

In medicine, this kind of effect doesn’t really matter. Someone’s reaction to a specific type of surgery doesn’t really change if a bunch of other people suddenly get access to that same type of surgery. But in social science, where we study behavior of people, this kind of effect can matter a lot⁴. This is because people in the world react to policy changes around them.

Maybe an experiment shows that increasing the amount of video in News Feed improves time spent on Facebook. Let’s say that it’s because most of the videos posted to Facebook are actually pretty good and interesting videos, and people like good, interesting content. You roll out the change to 100% of users. Media companies and influencers realize that if they post a video, it’s going to get boosted in feed. Suddenly, tons of content is being produced in video form. A lot of it is actually kind of terrible - it would be better as a news article or a photo. User welfare is actually worse than before, and people start using Facebook less because they’re sick of all the crappy video content.

One way to think about it is that by rolling out the change, you change the “game” (in the game theoretic sense) of the environment, and that the natural equilibrium actions of the actors within your environment will also change. When you run a randomized experiment, it’s really hard to structure it such that the “policy environment” equilibirum shifts at the same time, because by definition you’re not actually changing the policy environment for everyone.

View Comments