Correlation is not causation is the most repeated sentence in popular science and among the least useful, because reciting it does not tell you what else might be true.
The useful version is a list. When two things are found together, here are the explanations always sitting on the table.
The three alternatives
One. The arrow runs the other way. Children who read more have better vocabularies. Reading may build vocabulary; having a vocabulary may make reading enjoyable enough to do more of. Both are plausible, and both are probably true at once.
Two. Something else causes both. The classic confounder. Families who do one thing tend to do many other things, and to differ in income, education, housing, health and time. Any of those could produce both sides of the association.
Three. Selection. The association may be created by who ended up in the sample. If participation, or survival in a longitudinal study, is related to both variables, a relationship can appear that does not exist in the population.
A fourth deserves mention: chance, particularly in small samples, as described in sample size in plain words.
Why controlling for things is not enough
Researchers know about confounding and adjust for it statistically, controlling for income, parental education and so on.
This helps and it does not solve the problem, for a reason worth stating plainly: you can only control for what you measured. A confounder nobody thought of, or thought of and could not measure, remains entirely uncontrolled, and the adjusted result looks reassuringly precise regardless.
Adjustment is also imperfect even for measured variables. Parental education recorded as a category is a crude proxy for everything it stands in for, and residual confounding survives.
So a paper reporting an association that persists after adjustment has shown something real about the data and has not established causation. The phrase after controlling for should raise your confidence somewhat and not transform it.
Designs that get closer
Several approaches do better than plain observation without being experiments.
Longitudinal designs measure the presumed cause before the outcome, which rules out the simplest reverse causation. Direction in time is necessary for causation and not sufficient.
Within person designs compare the same individual at different times, removing everything stable about them. Applied to screen time research, these have generally produced smaller effects than comparisons between people, which is informative.
Sibling and twin designs compare children within the same family, removing shared family factors. These are powerful and have repeatedly shrunk associations that looked substantial in the general population.
Natural experiments use something arbitrary, such as a policy change or a date cut off, to create groups that differ in one respect for reasons unrelated to family characteristics. The relative age effect discussed in what school readiness means is studied this way.
Trials, and why there are so few
Random assignment solves confounding at a stroke, because chance determines who gets what and groups differ only by accident.
In research on children, most interesting questions cannot be randomised. You cannot assign a home language, a screen habit, a family structure or a temperament. What can be randomised is an intervention, and intervention trials are where education research has concentrated, including much of the work funded by the Education Endowment Foundation.
Trials have their own limits. They test what happens when a programme is introduced into particular schools for a particular period, which is not the same as the underlying question. And a trial finding nothing is often reported as evidence that the underlying idea is wrong, when it may be evidence that this implementation, at this dose, in these settings, did nothing.
The language that hides the gap
Coverage signals the distinction with vocabulary, and it is worth learning to hear.
Associated with, linked to, more likely to and predicts are correlational. Causes, leads to, results in and improves are causal. The transition usually happens between the paper and the headline, and nobody involved has to lie for it to occur.
Words like risk and protects against sound causal and are frequently used correlationally. Boosts, harms and damages are almost always overreach when applied to observational findings about children.
Why this is the most useful thing in the section
Because nearly every claim a parent encounters is observational. Nobody randomly assigns bedtimes, reading habits, nursery attendance or family structure. The evidence available on almost every question that matters to a household is correlational, and it will remain so.
That does not make it worthless. It makes it evidence of a particular kind, capable of bearing a particular amount of weight. Knowing how much is the skill.
The habit worth building is small: when you meet an association, spend ten seconds generating the reverse explanation and the common cause explanation before accepting the obvious one. It is usually easy, and it is usually enough.
For the wider frame, see what one study can tell you.