How to Ask a Question in Psychological Science

Does social media harm adolescent wellbeing?

Orben (2019) found a negative association between social media use and wellbeing among British teenagers. The size of that association was comparable to the effect of wearing glasses.

Courts are deciding right now

On 18 February 2026, CNN reported testimony in ongoing US litigation over adolescent social media use. Courts, legislators, and parents are making consequential decisions on the basis of psychological evidence.

Whether that evidence supports causal conclusions depends on measurement, design, and the target population.

Association \neq causation

How do we move from an observed association to a defensible causal claim?

This course builds five components: a well-defined intervention, valid measures, a confounding-control strategy, a clearly specified target population, and sensitivity analyses.

Course throughline

• Measurement is causal: responses are effects of constructs and of other causes.
• Three bias families recur: confounding bias, selection bias, and measurement-error bias.
• Causal inference provides workflows and sensitivity analyses to diagnose bias and quantify uncertainty.
• Comparative research asks which subpopulations are affected differently, and with what uncertainty.

Three problems

Measurement: what are we capturing?

A construct is the abstract attribute (e.g., wellbeing). An item is a single question (“How satisfied are you with your life?”). A scale combines items into an overall score (e.g., the WHO-5).

Before we can ask whether social media harms wellbeing, we need to measure both.

“Social media use” is not one thing

A single item (“How many hours per day?”) collapses these into one number.

Five types of measurement validity

Three uses of “validity”

The same word, three different meanings. Measurement validity is a precondition for the other two.

Three structural problems

Even with perfect measurement, an observed association can be spurious. The next three slides introduce three structures that generate non-causal associations. Learning to recognise them is the first step toward causal reasoning.

Reverse causation

\text{Social media} \longrightarrow \text{Wellbeing} \quad \text{or} \quad \text{Wellbeing} \longrightarrow \text{Social media?}

If unhappy teenagers turn to social media for distraction, the arrow runs from low wellbeing to high usage. The association is real, but the causal direction is wrong.

Common causes

A confounder is a common cause of both the exposure and the outcome. If the common cause is unmeasured, the spurious association cannot be removed by adjustment.

Collider bias

A collider is a common consequence of two variables. Unlike a confounder, it creates bias only when we condition on it. Restricting the sample to high achievers opens a spurious path.

Collider demo (R)

set.seed(123)
n <- 100000
h <- rnorm(n)
s <- rnorm(n)
a <- 0.6 * h + 0.6 * s + rnorm(n, sd = 0.6)

high_a <- a > quantile(a, 0.75)

c(
  full_sample = cor(h, s),
  conditioned_on_high_a = cor(h[high_a], s[high_a])
)

          full_sample conditioned_on_high_a 
         -0.002659378          -0.339749286 

Near zero in the full sample. Strong negative association after conditioning on high academic performance.

Internal validity

Does the study design support a causal claim, free of confounding?

If income confounds the social media–wellbeing association and is not adjusted for, the study lacks internal validity. The problem is structural, not statistical: no amount of data changes it.

External validity

Does the causal claim generalise beyond the study sample?

Orben’s data came from one country, one cohort, one time period (UK, pre-2019). Platforms, algorithms, and adolescent usage patterns have all changed since.

Measurement \times generalisation

A scale that operates differently across groups turns a measurement problem into a generalisation problem. If a wellbeing scale has different factor structures in different populations, apparent group differences may reflect measurement artefact, not substantive variation.

What do we need?

Next week

Causal diagrams give us a formal language for structural assumptions. The five elementary structures (independence, causation, fork, chain, collider) are the building blocks for every confounding-control argument in this course.

Readings

Required and optional readings for each week are listed on the course readings page.