The effect of deactivating Facebook and Instagram on users' emotional state

From the paper PDF (https://www.nber.org/system/files/working_papers/w33697/w336...):

> We estimate that users in the Facebook deactivation group reported a 0.060 standard deviation improvement in an index of happiness, anxiety, and depression, relative to control users. The effect is statistically distinguishable from zero at the p < 0.01 level, both when considered individually and after adjusting for multiple hypothesis testing along with the full set of political outcomes considered in Allcott et al. (2024). Non-preregistered subgroup analyses suggest larger effects of Facebook on people over 35, undecided voters, and people without a college degree.

> We estimate that users in the Instagram deactivation group reported a 0.041 standard deviation improvement in the emotional state index relative to control. The effect is statistically distinguishable from zero at the p = 0.016 level when considered individually, and at the p = 0.14 level after adjusting for multiple hypothesis testing along with the outcomes in Allcott et al. (2024). The latter estimate does not meet our pre-registered p = 0.05 significance threshold. Substitution analyses imply this improvement is achieved without shifts to offline activities. Non-preregistered subgroup analyses suggest larger effects of Instagram on women aged 1824.

Perhaps it wasn't clear what I meant. When I said significantly, I meant it in the colloquial sense, not in the statistical significance sense.

I was looking for a more digestable figure describing the extent of improvements, not whether the study found them confidently distinguishable (which I just assumed they did based on the wording, good to know they didn't for Instagram).

A 0.060 standard deviation improvement is super small. If the average person rates their happiness/anxiety/depression score at, say, 50 out of 100, and the standard deviation (how spread out people’s scores are) is around 10 points, then 0.060 SD = 0.6 points. So quitting Facebook gave an average person a ~1% bump in mood score. Instagram was even smaller: ~0.4 points, or 0.8%.

It's real, but barely noticeable for most people—unless you're in a more affected subgroup (e.g. undecided voters or younger women). Your experience feeling way better likely means you were an outlier (in a good way).

On what scale? What do 'points' on the scale mean? Without knowing those things, we can't say what 6 or 60 points mean.

This is what I was interested in, thank you!

So, ELI5 level.

People who use Facebook also may feel depression, from very strong to none at all. In the middle of this interval there's the "expected value" point, sort of an average level of feeling depressed. This point is at an equal distance from the "most depressed users" group, and from the "not depressed at all" group. Let's call this distance of depression strength a "standard deviation".

Now, the users who stopped using Facebook became slightly less depressed, by 6% of that "standard deviation" range. If you buy a small coke at a McDonald's, then take one sip, you make it about 6% smaller. It's not unnoticeable (you've made that refreshing sip), but about 15 more such sips still remain!

In other words, there is an effect which can definitely be noticed ("statistically significant"), but it's not a big deal either.

The best thing you can do is compare it to another study, since turning 0.06 standard deviations into a percentage of happiness isn’t going to be that telling.

In general, 0.2 is considered a small effect. So 0.06 is quite small — likely not a practically noticeable change in well-being. But impressive to me when I compare it to effect sizes of therapy interventions which can lie around 0.3 for 12 weeks.

Quote:

> “50 randomized controlled trials that were published in 51 articles between 1998 and August 2018. We found standardized mean differences of Hedges’ g = 0.34 for subjective well-being, Hedges’ g = 0.39 for psychological well-being, indicating small to moderate effects, and Hedges’ g = 0.29 for depression, and Hedges’ g = 0.35 for anxiety and stress, indicating small effects.”

(Source: The efficacy of multi-component positive psychology interventions, 2019 — https://www.researchgate.net/publication/331028589_The_Effic...)

On the contrary, reporting changes relative to the standard deviation of a control group frees you from scales and their meanings, because it relates the observed change to the normal spread of scores before the intervention. In this way, you don't need to know the scale and its meaning to know if a change is big or small, and from a statistical perspective, that's (almost) all you need to find if a change is significant or due to random chance. Of course, looking back at the original scale and its meaning can help interpreting the meaning of the results in other ways

This is a very useful insight, thank you. Wouldn't have occurred to me to check something like that.

Multiply by .37 to get PERCENTILE change, not percent.

If you were average happiness, and you improve that by 1 stdev, you are now happier than 87% of your peers (when you were at 50%ile before). 0.6 stdev improvement would be vs 72% of your peers.

So to put it colloquially, if you have 7 friends, and you were in the middle of them (4th happiest), by quitting Facebook you are now happier than all but 1 one them.

Multiply by .37 to get PERCENTILE ranking change, not percent. If you were average happiness, and you improve that by 1 stdev, you are now happier than 87% of your peers (when you were at 50%ile before). 0.6 stdev improvement would be vs 72% of your peers.

So to put it colloquially, if you have 4 friends, and you were in the middle of them (3rd happiest aka happier than 2 of them), by quitting Facebook you are now happier than all but 1 one them (aka happier than 3 of them).

AKA for every 4 friends you have you can jump ahead of 1 of them in the happiness race by quitting facebook.

Standard deviation helps, but you still need to know: standard deviation of what? It's no different than saying someone scored 78% - 78% of what? What is it in the denominator? Also, different scales can represent the same thing differently.

Secondly, the impact of the difference isn't known - you don't know the curve representing the relationship of score to impact. In some contexts a little change is meaningless - the curve is flat; in others the curve is steep and it can be transformational. And impacts only sometimes scale linearly with performance or score, of course.

Without that knowledge, standard deviation means nothing beyond how unusual, in the given population, the subject's performance is.

Caveat: I'm not very smart

So, if 0.3 is 12 weeks of therapy, then 0.06 is ~2.5 weeks of therapy (0.3/0.06 = 2.4), assuming you pick any 2.5 random weeks of the 12 week course.

Yes, I'm sure the first session is the most important and then a logarithmic curve of blah blah blah.

Essentially, deleting FB is not much, but it's not nothing either.