Causality is not whether two things go together. It is whether one produces the other, and how sure we can be.
Almost every weak answer about causality ends in the same place: correlation does not equal causation, therefore we cannot conclude anything. That sentence is a retreat, not an evaluation, and it stops exactly where the marks begin.
Three things that are always true of it
It is a ladder. Causal claims come in strengths. The error is not making a causal claim. It is standing on a higher rung than your evidence reaches. So the question is always which rung this design reaches, and which rung the writer is standing on.
It is probabilistic. Causes shift the odds, they do not decide. One counterexample does not refute a cause. Smoking causes lung cancer even though some smokers never develop it.
It needs a pathway. A mechanism answers how X could produce Y. Without one you have a claim, not an explanation.
The ladder
Six rungs. Every one is a real claim worth something.
| Rung | The claim | What it still cannot rule out |
|---|---|---|
| 1 | They go together | Chance, reverse direction, a confounder |
| 2 | One predicts the other | Says nothing about why. Changing X may change nothing |
| 3 | One came first | Direction addressed, confounding untouched, attrition still bites |
| 4 | Alternatives controlled | Known confounders handled, unknown ones remain |
| 5 | Changing it changes the outcome | Unknown confounders spread by randomisation, but a weak manipulation or dropout still beats it |
| 6 | And we know why, and it replicates | Pathway specified, dose-response, converging designs |
Rung 5 needs manipulation with random allocation. Rung 6 needs mechanism plus independent replication.
Placing a study on this ladder, in one sentence, is worth more than a paragraph of general caution.
What else fits?
When you find a causal claim, the productive move is not to doubt it. It is to generate specific alternatives.
Seven kinds are worth having ready. Reverse direction. Bidirectional influence. Confounding by a third variable. Genetic confounding in family findings. Selection into groups. Regression to the mean when extreme scorers were chosen. And plain chance, especially in a small study.
Three specific alternatives beat one vague mention of a third variable. And after generating them, pick the most plausible one and say why. That choice is the analysis.
Regression to the mean, because it catches everyone
If a group was selected because they scored at an extreme, they will drift toward average on retesting even if nothing was done to them.
This is why a support programme given to the most anxious students will show improvement whether or not it works, and why any study that selected on an extreme score needs a control group before its result means anything.
It is one of the most useful things you can know, because it explains a huge number of apparent effects and almost nobody in an exam hall mentions it.
Two words people mix up
A mediator is a step in the chain. X changes M, M changes Y. It answers how.
A moderator changes the strength or direction of a relationship. It sits outside the chain and adjusts the dial. It answers for whom and when.
Mediator, how. Moderator, when. Learn it that way and you will stop mixing them up.
Two things not to say
A predictor is not something to fix. Postcode may predict dropout. Changing someone's postcode does not change what it was standing in for. Confusing a proxy with a cause is a specific, nameable error and spotting it is worth marks.
Multiple causes do not make causality unstudiable. "Behaviour has many causes so we cannot say" is a retreat. Say which cause this design isolates, and what the rest would need.
How to write about it
Six moves, in order.
Name the rung this design reaches. Give three alternatives, not "a third variable". Pick one and say why it is most plausible here. Propose a mechanism for how this could work. Name the design that would separate the explanations. State what survives, meaning the strongest claim still standing.
The last one matters most. Strong answers do not say causation cannot be established. They say what has been established, and name the design that would establish the rest.
With no control group, the improvement is equally consistent with the intervention, with regression to the mean given that the highest scorers were selected, and with ordinary development over a school term. A randomised comparison group would separate the three. What the study does establish is that scores rose over eight weeks in this group, which is worth knowing and is not evidence that the programme caused it.
Where causality meets the other concepts
Two pairings produce the sharpest lines available in an exam.
Causality and measurement. If the outcome is measured badly, no amount of experimental control saves the causal claim. You have cleanly manipulated one thing and precisely measured something else. The design supports a causal claim about scores on this task, but the task is a narrow version of the construct, so the claim about the construct is weaker.
Causality and perspective. Different levels of analysis find different causes of the same behaviour without contradicting each other. The biological and sociocultural accounts are often not competing. They describe different points in the same causal system.
That second one is a genuinely sophisticated thing to write, and it takes one sentence.
Where this gets examined
Paper 1 Section C can name causality as the concept for any Area of Study.
Paper 2 Section B rewards it whenever an unseen study measured variables without manipulating anything, which is most of the time.
Paper 3 questions 1, 2 and 4 all turn on it, usually as the gap between what a design supports and what the researchers concluded.
And in your IA, choosing a method means choosing what kind of causal claim you could make, which is one of the strongest justifications available in Criterion B.
Can you do this yet?
- Name the six rungs and place a study on one
- Give three specific alternative explanations instead of saying "third variable"
- Explain why one counterexample does not disprove a probabilistic cause
- Tell mediator from moderator using the how and when test
- Explain why randomisation handles confounders nobody thought of
- Explain how regression to the mean can fake a treatment effect
- Explain the difference between a predictor and something worth changing
- Explain why biological and social accounts are not necessarily rivals
- Name the design that would move a study one rung up
- State the strongest claim a design supports, instead of saying causation cannot be established
That last one is the whole skill. If you can do nothing else on this list, do that.
If you want the concept worked through with labs and worked answers, that is what the PsychExcel Six Concepts Workbook is built for.