Psychology › Section B · Psychological research › Topic B.3

Analyzing findings: correlation and experiments

Children with bigger feet read better - but feet do not teach reading. This topic shows how psychology's data become knowledge: what correlations can and cannot say, how experiments earn the word "cause," and how the field checks itself through peer review, replication, reliability and validity.

The correlation coefficient

When two variables change together, they are correlated, and the correlation coefficient - usually written r - measures the relationship on a scale from -1 to +1. The number's size is the strength: the closer to 1 or -1, the more predictable one variable is from the other, so -0.9 is as strong as +0.9 and far stronger than 0.3; unrelated variables sit near 0. The number's sign is the direction: positive correlations move together (more of one, more of the other); negative correlations move oppositely (more of one, less of the other). Correlational research is what surveys, archival data and most observational studies produce: relationships, discovered and measured.

Correlation is not causation

A correlation tells us variables travel together; it does not tell us why. Sometimes one variable does cause the other - but a third factor, a confounding variable, may be moving them both. Among primary-school children, shoe size and reading score rise together, and the confounder is age: older children have bigger feet and more years of reading practice. Even with no visible confounder, a correlation alone does not establish cause. The link between smoking and lung cancer began as correlations; the US Surgeon General's 1964 report judged it causal only after many kinds of evidence converged. A correlation starts an investigation; it does not end one.

Illusory correlations and confirmation bias

Many people with arthritis are sure their pain tracks the weather. When Donald Redelmeier and Amos Tversky (1996) compared 18 patients' pain ratings with daily weather records over 15 months, they found no meaningful link, even though most patients believed in one. Such beliefs survive through illusory correlation - seeing a relationship that is not there - by noticing the matches and forgetting the misses. Its engine is usually confirmation bias - seeking evidence for a hunch while ignoring contrary evidence - plus the availability of vivid memories. The remedy this course keeps prescribing: count, do not recall.

Designing an experiment

To test a hypothesis about cause, we cannot rely on observations: we conduct an experiment. The basic design compares two groups - the experimental group and the control group - identical except for one thing: the experimental manipulation, the treatment being tested. Every experiment names two variables: the independent variable is what the experimenter manipulates (say, self-testing versus rereading), and the dependent variable is what is measured to detect the effect (scores on a later quiz) - it DEPENDS on the manipulation.

Random sampling and random assignment

Two different random acts, easily confused. A random sample is drawn from the population - every member has an equal chance of selection - so a large enough sample tends to mirror the population and results generalise: a state cannot test every one of its students, so a few hundred are drawn at random. Random assignment happens INSIDE the experiment: chance decides who joins the experimental group and who the control, spreading pre-existing differences evenly. Sampling earns the right to generalise; assignment earns the right to say because.

Placebos and blind designs

Expectations change outcomes: participants who believe they received a treatment often improve anyway - the placebo effect - and researchers who know who got it may score favourably - experimenter bias. The defences are blind designs: single-blind (participants do not know their group) and double-blind (neither participants nor the researchers working with them know), which controls both expectation effects at once. Hence the placebo control - an inert treatment that looks identical - so the treatment is compared against expectation, not against nothing. After the data are in, a statistical test asks whether the group difference is significant: by convention, less than a 5% chance of a difference this large if the groups did not truly differ.

Worked example: questions experiments cannot ask

Some variables cannot be assigned. You cannot randomly assign a participant's sex, age or childhood experiences - so studies comparing pre-existing groups are quasi-experimental. 1. Question: do only children have better vocabularies than children with siblings? 2. Problem: family size cannot be assigned, and the groups may also differ in parents' income or age. 3. Verdict: the study can describe a difference but cannot prove that siblings cause it. Ethics fences the territory further: testing whether severe childhood stress harms adult health by experiment would mean assigning children to harm - unthinkable. Such questions rely on correlational designs, carefully hedged.

Worked example: designing a study-method experiment

Hypothesis: testing yourself on a passage beats rereading it. 1. Sample: 120 volunteers from a college's first-year students (results generalise only to similar students unless the sample is random). 2. Random assignment: 60 practise recalling the passage (experimental group), 60 reread it (control). 3. Independent variable: study method. Dependent variable: score on a quiz one week later. 4. Blinding: quiz markers do not know who was in which group. 5. Reading: if the self-testing group scores significantly higher, the method is the likeliest cause, since it was the one planned difference. 6. Check: a different room or time of day for one group would be a confound; identical conditions and random assignment protect against it. (Real studies of this kind, such as Roediger and Karpicke, 2006, have found a self-testing advantage.)

Peer review and replication

Before a study joins the field's knowledge, it faces peer review: experts screen submitted manuscripts for sound design, analysis and honest reporting, and journal editors publish, request revisions, or reject. The screen is imperfect: Andrew Wakefield's 1998 paper claiming a vaccine-autism link passed review, was later shown to be fraudulent, and was retracted in 2010. Then comes replication: other scientists repeat the study with new samples. In 2015 the Open Science Collaboration repeated 100 published psychology studies and obtained statistically significant results in only about 36% of them - a centrepiece of the "replication crisis" that has pushed the field towards larger samples and pre-registered plans.

Reliability and validity

Any measurement must be judged on two axes. Reliability: consistency - the tool produces the same result across uses: inter-rater reliability (different observers agree), internal consistency (survey items measuring the same thing agree), test-retest reliability (the same measure, repeated, gives the same answer). But consistency is not correctness: a miscalibrated kitchen scale reliably reports the wrong weight every time. Validity is correctness: the tool measures what it claims to measure. Imagine a "creativity test" that only asks people to define hard words: scores could be highly consistent from week to week, yet they would mostly measure vocabulary, not creativity - reliable but not valid.

Common mistakes to avoid

First, reading causation into correlation - a confounder may be hiding (age behind shoe size and reading). Second, trusting remembered matches over counts, as with weather and arthritis pain. Third, confusing random sampling (generalise) with random assignment (cause): they earn different claims. Fourth, forgetting the dependent variable depends: name what is manipulated and what is measured, every time. Fifth, assuming quasi-experimental group differences are causal: pre-existing groups differ in unassignable ways. Sixth, mistaking reliability for validity: a consistent scale can be consistently wrong. Seventh, treating a single significant result as settled: many have failed to replicate.

Key takeaways

Correlational research measures relationships (r, -1 to +1) but cannot establish cause on its own: confounders may drive both variables (age behind shoe size and reading), and illusory correlations survive on confirmation bias (weather and arthritis pain, Redelmeier and Tversky, 1996). Experiments earn the word "because": groups differ only in the independent variable, its effect read on the dependent variable; random samples generalise, random assignment isolates cause; placebos and blind designs neutralise expectation; significance tests judge chance. Quasi-experiments and ethics limit causal claims. Peer review is imperfect (Wakefield, retracted 2010) and replication is the stronger check (about 36% in 2015). Reliability is consistency, validity is correctness.

Flashcards

Flashcard

Correlation coefficient (r)

-1 to +1: strength and direction of a relationship

-0.9 is as strong as +0.9; both beat 0.3.

Test your understanding

Question 1 of 5

Among primary-school children, shoe size and reading score are positively correlated. The best explanation is:

Question 2 of 5

An experiment compares self-testing with rereading and measures quiz scores a week later. What is the dependent variable?

Question 3 of 5

Random assignment of participants to groups is what allows a study to:

Question 4 of 5

A double-blind study means:

Question 5 of 5

A "creativity test" gives each person nearly the same score every week, but the scores mostly reflect vocabulary. The test is:

← Previous topicNext topic →