Your key claim here is that the effects are "tiny". I think you're quite wrong on this.
I haven't yet modeled the RCT data, but the r=0.17 correlations found in the observational literature, if real and causal, can fully account for the 50% increase in teen girl depression. I was able to model this using maximally simple assumptions. Read this short blog post to understand how.
Or alternatively, consider the effects of childhood lead exposure, which you yourself have described as a major issue. The correlation of exposure to adult IQ is 0.11, which is smaller than the social media / mental health correlations https://www.judiciary.senate.gov/imo/media/doc/Haidt%20Testimony.pdf
Full disclosure: I think there's only a 50% chance social media is causing major challenges for teen mental health. In this comment, I just want to narrowly challenge your claims about "trivially" small effect sizes.
I always appreciate substantive comments. Chris Said’s model suggests that if social media is maximally assumed to be the only driver of teen girls’ depression (which he acknowledges it is not), even a small correlation between social media use and girls’ depression (which he estimates at Cohen’s d=+0.17, itself a maximum assumption based on a small, selected set of studies) is sufficient to explain the increase from 13% to 19% in girls’ depressive episodes after 2007.
Well, yes… in the most rarified hypothetical sense. A Cohen’s d-value of +0.17 means an association between potential cause (social media use) and effect (mental health) for 7% of the population, more than the 6% absolute increase in girls’ depression – again, assuming social media is the sole factor.
I calculate from the 2021 CDC survey that the association between poor mental health for girls who use screens the most (5+ hours/day) versus the least (<1 hour/day) is actually a bit larger than Said, Haidt, and others suggest, d = +0.30 – but still small, an effect on 12% of all girls, at most.
But then, problems mount. Let me state this politely: Studies asserting that social media damages teenage mental health that do not include a variable controlling for parental abuses (or a reasonable surrogate) are invalid on their face. It’s like talking about health problems in 2020 without mentioning COVID.
This is how big parental abuse is: analysis of the CDC survey yields a d=+0.87 for its association with girls’ poor mental health, indicating some 80% are negatively affected – not 7%, as for social media.
The complications get worse. Girls who use screens the most are also the most likely to have been abused by parents (d=+0.24). So, it is likely that nearly all the association between screen use and girls’ depression is spurious, the result of ignoring abuse’s (and possibly other factors') big effect on girls’ depression.
And still more: the association between girls’ worst outcome, attempted suicide, and parental abuse is staggering, d=+0.97 – but MINUS 0.16 for attempted suicide versus screen time (yes, girls who use screens the most are LESS likely to attempt suicide).
The bottom line I argue the relevant studies show: abused, depressed girls are more likely to use social media, which may slightly reduce their odds of actual harm.
Did parental abuses increase as girls’ depression increased? We don’t know. We haven’t cared enough about adults’ abuse of teenagers to install consistent measures. But we know parent-age drug/alcohol abuse sure did – from 483,000 (2007) hospital ER cases among ages 25-64 to 1,472,000 (2021) – only the tip of the overdose iceberg. More addicted adults could indicate more domestic abuse, or it could depress girls in and of itself, or both.
Hopefully, the more precise questions in the CDC’s 2023 survey will allow better study of these issues. Until then, we should adopt a much humbler stance commensurate with our ignorance and halt efforts to ban or curb teens’ screen use.
Also, my childhood lead exposure correlation was with youth violent crime and gun homicide rates over the last 30 years, which yield very high d-values as well (d=0.88 to 0.92). Policy changes affecting entire populations should be based on high associations like these, not tiny ones like d=0.17, which indicates individual, not mass, measures.
The key word here is *IF* real and causal, that is. And that is a VERY big "IF", of course. And even then, that is only because social media went from nil in the early 2000s to the the vast majority of the population in a short period of time. Small effect sizes like this should be approached with caution in terms of drawing causal inferences, as they are very likely to be due to chance, bias, or residual or unmeasured confounding.
(At least with lead exposure and IQ, that has fulfilled virtually all of the Bradford-Hill criteria of causation.)
Researcher 1: "MY effect size is bigger than YOUR effect size!"
Researcher 2: "Oh yeah, well MY p-value is bigger than YOUR p-value!"
Researcher 1: "Yes, we know." (That's not a good thing, by the way.)
"But...but...precautionary principle! Prevention paradox! Collective action problem! Or [insert whatever disingenuous buzzword here]."
Your key claim here is that the effects are "tiny". I think you're quite wrong on this.
I haven't yet modeled the RCT data, but the r=0.17 correlations found in the observational literature, if real and causal, can fully account for the 50% increase in teen girl depression. I was able to model this using maximally simple assumptions. Read this short blog post to understand how.
https://chris-said.io/2022/05/10/social-media-and-teen-depression/
Or alternatively, consider the effects of childhood lead exposure, which you yourself have described as a major issue. The correlation of exposure to adult IQ is 0.11, which is smaller than the social media / mental health correlations https://www.judiciary.senate.gov/imo/media/doc/Haidt%20Testimony.pdf
Full disclosure: I think there's only a 50% chance social media is causing major challenges for teen mental health. In this comment, I just want to narrowly challenge your claims about "trivially" small effect sizes.
I always appreciate substantive comments. Chris Said’s model suggests that if social media is maximally assumed to be the only driver of teen girls’ depression (which he acknowledges it is not), even a small correlation between social media use and girls’ depression (which he estimates at Cohen’s d=+0.17, itself a maximum assumption based on a small, selected set of studies) is sufficient to explain the increase from 13% to 19% in girls’ depressive episodes after 2007.
Well, yes… in the most rarified hypothetical sense. A Cohen’s d-value of +0.17 means an association between potential cause (social media use) and effect (mental health) for 7% of the population, more than the 6% absolute increase in girls’ depression – again, assuming social media is the sole factor.
I calculate from the 2021 CDC survey that the association between poor mental health for girls who use screens the most (5+ hours/day) versus the least (<1 hour/day) is actually a bit larger than Said, Haidt, and others suggest, d = +0.30 – but still small, an effect on 12% of all girls, at most.
But then, problems mount. Let me state this politely: Studies asserting that social media damages teenage mental health that do not include a variable controlling for parental abuses (or a reasonable surrogate) are invalid on their face. It’s like talking about health problems in 2020 without mentioning COVID.
This is how big parental abuse is: analysis of the CDC survey yields a d=+0.87 for its association with girls’ poor mental health, indicating some 80% are negatively affected – not 7%, as for social media.
The complications get worse. Girls who use screens the most are also the most likely to have been abused by parents (d=+0.24). So, it is likely that nearly all the association between screen use and girls’ depression is spurious, the result of ignoring abuse’s (and possibly other factors') big effect on girls’ depression.
And still more: the association between girls’ worst outcome, attempted suicide, and parental abuse is staggering, d=+0.97 – but MINUS 0.16 for attempted suicide versus screen time (yes, girls who use screens the most are LESS likely to attempt suicide).
The bottom line I argue the relevant studies show: abused, depressed girls are more likely to use social media, which may slightly reduce their odds of actual harm.
Did parental abuses increase as girls’ depression increased? We don’t know. We haven’t cared enough about adults’ abuse of teenagers to install consistent measures. But we know parent-age drug/alcohol abuse sure did – from 483,000 (2007) hospital ER cases among ages 25-64 to 1,472,000 (2021) – only the tip of the overdose iceberg. More addicted adults could indicate more domestic abuse, or it could depress girls in and of itself, or both.
Hopefully, the more precise questions in the CDC’s 2023 survey will allow better study of these issues. Until then, we should adopt a much humbler stance commensurate with our ignorance and halt efforts to ban or curb teens’ screen use.
Also, my childhood lead exposure correlation was with youth violent crime and gun homicide rates over the last 30 years, which yield very high d-values as well (d=0.88 to 0.92). Policy changes affecting entire populations should be based on high associations like these, not tiny ones like d=0.17, which indicates individual, not mass, measures.
Well-said. There is no good reason to throw out the proverbial baby with the bathwater based on such a small effect size of questionable causality.
BINGO
The key word here is *IF* real and causal, that is. And that is a VERY big "IF", of course. And even then, that is only because social media went from nil in the early 2000s to the the vast majority of the population in a short period of time. Small effect sizes like this should be approached with caution in terms of drawing causal inferences, as they are very likely to be due to chance, bias, or residual or unmeasured confounding.
(At least with lead exposure and IQ, that has fulfilled virtually all of the Bradford-Hill criteria of causation.)
Well-said as usual, Mike! Keep up the great work 😊