I suspect social media’s negative effects occur at the network level, and thus cannot be detected by these studies, which are all conducted at the individual level.
There are many ways the network effects could kick in, but the simplest is easy enough to explain. Social media makes you unfun and unavailable to others in the real world. This doesn’t make you unhappy. It makes everyone else unhappy. https://chris-said.io/2021/08/14/teens-loneliness-social-media/
> Since this a longitudinal study, people serve as their own control, which is good enough probably. This way you control for any stable factors for persons, but any factors that changed between the two weeks of study would have messed up results. I mean, if in the week of no social media, nothing extraordinary happens, and in the follow-up back-to-normal week, a meteor strikes earth and kills a billion people, then comparing the two weeks of results would be a very bad idea, despite it being the same people. OK, so anyway (no meteors), they find that ... nothing on emotional effects (positive or negative effect), but people are somewhat more bored and crave social media for that week of absence.
To me, this doesn't really address the conceptual problem with using longitudinal studies for causal inference. Instead, I think the problem is this:
Why don't simple correlation work for establishing causality cross-sectionally? Mostly due to the possibility of confounders. That is, there are presumably some underlying reasons why people vary between each other in their amount of social media use, and these underlying reasons might also directly affect mental health, rather than *just* affecting social media use, leading to non-causal correlations.
But there's nothing special about cross-sectional data for this. Longitudinally, there are presumably also some reasons that people vary from moment to moment in social media use, and those reasons could *also* directly affect mental health, rather than just affecting social media use. So longitudinal data doesn't solve anything.
Well, that's not true, there are two things that longitudinal data does solve:
1. If there genuinely is a causal connection, then *any* way of splitting the variance up into components should still preserve the correlation. Whether that be between-person vs between-time, genetic vs nurture vs other stuff, between-group vs within-group, etc.. So looking at these sorts of variance components can give you different perspective on how stable the effect is.
2. If you believe one variance component is more likely to be confounded than another, then this sort of data can throw out the confounded component. But it's not immediately obvious that the (presumably-mostly-genetic) differences between people are more confounded than the (presumably-mostly-environmental) differences within people when it comes to social media use and mental health.
I suspect social media’s negative effects occur at the network level, and thus cannot be detected by these studies, which are all conducted at the individual level.
There are many ways the network effects could kick in, but the simplest is easy enough to explain. Social media makes you unfun and unavailable to others in the real world. This doesn’t make you unhappy. It makes everyone else unhappy. https://chris-said.io/2021/08/14/teens-loneliness-social-media/
> Since this a longitudinal study, people serve as their own control, which is good enough probably. This way you control for any stable factors for persons, but any factors that changed between the two weeks of study would have messed up results. I mean, if in the week of no social media, nothing extraordinary happens, and in the follow-up back-to-normal week, a meteor strikes earth and kills a billion people, then comparing the two weeks of results would be a very bad idea, despite it being the same people. OK, so anyway (no meteors), they find that ... nothing on emotional effects (positive or negative effect), but people are somewhat more bored and crave social media for that week of absence.
To me, this doesn't really address the conceptual problem with using longitudinal studies for causal inference. Instead, I think the problem is this:
Why don't simple correlation work for establishing causality cross-sectionally? Mostly due to the possibility of confounders. That is, there are presumably some underlying reasons why people vary between each other in their amount of social media use, and these underlying reasons might also directly affect mental health, rather than *just* affecting social media use, leading to non-causal correlations.
But there's nothing special about cross-sectional data for this. Longitudinally, there are presumably also some reasons that people vary from moment to moment in social media use, and those reasons could *also* directly affect mental health, rather than just affecting social media use. So longitudinal data doesn't solve anything.
Well, that's not true, there are two things that longitudinal data does solve:
1. If there genuinely is a causal connection, then *any* way of splitting the variance up into components should still preserve the correlation. Whether that be between-person vs between-time, genetic vs nurture vs other stuff, between-group vs within-group, etc.. So looking at these sorts of variance components can give you different perspective on how stable the effect is.
2. If you believe one variance component is more likely to be confounded than another, then this sort of data can throw out the confounded component. But it's not immediately obvious that the (presumably-mostly-genetic) differences between people are more confounded than the (presumably-mostly-environmental) differences within people when it comes to social media use and mental health.