Psychology as a Science & Research Methods

This is Unit One of Psychology, Part Two: measurement, and cause. Part One covered ways of looking at behavior. This part is about the difference between noticing that two things move together and being entitled to say one caused the other — which is the single most consequential distinction in the whole of research methods. And the course states it more strongly than most treatments do. Correlational data cannot be converted into causal data by any amount of further reasoning about the same data. Not by a better argument, not by a bigger sample, not by a more sophisticated analysis. The design decides what the data can support. So the part builds the design that does license a causal claim, piece by piece — a manipulated variable, a measured one, groups built by chance, and every competing explanation closed off deliberately. Then statistical significance, which is a technical term with a narrow meaning and three popular misreadings. Then the ethics that stand between a proposal and a study. And a capstone that runs one health headline backwards to what its data actually support. By the end you will be able to do five things. One: say why correlational data stays correlational. Two: read a correlation coefficient in three steps. Three: say what random assignment does that random sampling does not. Four: say what a confound costs. Five: say what significance does and does not establish.

Three entries, and the third one is stronger than the usual warning. It does not say be careful about inferring cause from correlation. It says the conversion cannot be done at all, on that data, by anyone.

The measurement.

Psych 0063 Correlation

A measured relationship in which changes in one variable accompany changes in another. Accompany is the whole content of the claim — nothing more is asserted.

The error.

Psych 0064 The Correlation–Causation Error

Treating a measured relationship between two variables as evidence that one causes the other. It is the most common error in reading research, by a wide margin.

And why it stays that way.

Psych 0065 Why Correlational Data Stays Correlational

Correlational data cannot be converted into causal data by any amount of additional reasoning about the same data. Not a caution about being careful — a statement about what that design can support.

Five entries built around one number and a three-step reading procedure — and the third step is the one people skip. Sign, then strength, and then what the number cannot establish no matter how large it is.

The number.

Psych 0067 The Correlation Coefficient

Running from minus one to plus one, with the sign giving direction and the absolute value giving strength. Two pieces of information in one number.

How to read it.

Psych 0069 Reading a Correlation Coefficient

Three steps — sign first, strength second, and then what it cannot establish. The third step is part of reading the number, not an afterthought about it.

The first case.

Psych 0070 Reading a Weak Positive Coefficient

Point one two is positive in direction and weak in strength, describing a relationship most individual cases will not follow. Real at the group level and unhelpful about any one person.

The second.

Psych 0072 Reading a Strong Negative Coefficient

Minus point six is negative and moderately strong — higher values on one variable accompanying lower values on the other. Still a relationship rather than a mechanism.

The third.

Psych 0074 Reading a Near-Zero Coefficient

Minus point zero three describes effectively no relationship in the sample measured. The sign is present and it carries no information at that magnitude.

Five entries on failure. The first two are about the data — a cause nobody measured, and a relationship that was never in the numbers. The last three are about the person reading, which is a harder problem because the person cannot see themselves doing it.

The unmeasured cause.

Psych 0076 The Third-Variable Problem

An unmeasured variable producing the relationship observed between two measured ones. Both measured variables are real, the relationship is real, and the explanation is somewhere else entirely.

The relationship that is not there.

Psych 0078 Illusory Correlation

A relationship a person perceives between two variables that the data do not show. The perception is genuine; the relationship is not.

The reader's bias.

Psych 0080 Confirmation Bias

Seeking and weighting evidence supporting a held belief while discounting evidence against it. It operates on both what gets looked for and what gets counted.

Its correction.

Psych 0081 Correcting for Confirmation Bias

Specifying in advance what evidence would count against the belief, before that evidence is collected. In advance is the entire mechanism — the same specification made afterwards does nothing.

And one more.

Psych 0083 Hindsight Bias

Seeing an outcome as having been predictable once that outcome is known. It makes past research look more obvious than it was, and past mistakes look more avoidable.

Four entries building the design that does license a causal claim. Manipulate one thing, hold the rest constant, measure the result — and then a twin entry separating two randomisations that sound alike and do entirely different jobs.

The design.

Psych 0085 The Experiment

Manipulating one variable, holding others constant, and measuring the effect — which is what licenses a causal claim. The licence comes from the design, not from the size of the result.

Two variables.

Psych 0087 Independent and Dependent Variables

The independent variable is the one manipulated; the dependent variable is the one measured for change. Naming which is which is naming the direction of the claim.

Building the groups.

Psych 0089 Random Assignment

Placing each participant into a condition by chance, so the groups differ only by chance at the start. Any difference at the end therefore has one candidate explanation.

And the confusion.

Psych 0090 Assignment Is Not Sampling

Random assignment builds the groups within a study. Random sampling builds the sample from a population. One protects the causal claim, the other protects the generalisation, and a study can have either without the other.

Five entries on what goes wrong and what keeps it out. The second one is precise in a way worth carrying — a confound does not make a result false. It makes the result unreadable, which is a different and in some ways worse outcome.

The problem.

Psych 0092 Confounding Variable

A variable varying alongside the manipulated one, making the cause of the result ambiguous. Two things changed, and the result cannot say which one did it.

What it costs.

Psych 0093 What a Confound Costs

A confound does not make a result false — it makes it unreadable, because two explanations fit the same data equally well. The finding may be entirely correct and there is no way to tell.

The basic control.

Psych 0095 Experimental and Control Groups

The experimental group receives the manipulation; the control group is treated identically except for it. Identically is a demanding word and the whole design rests on it.

One reason it is needed.

Psych 0097 The Placebo Effect

A measurable change produced by the expectation of treatment rather than by the treatment. Measurable is the point — expectation produces real effects, not imagined ones.

Closing that off.

Psych 0099 Single-Blind and Double-Blind Designs

In a single-blind design the participant does not know their condition; in a double-blind design neither the participant nor the experimenter does. The second exists because the experimenter's expectations leak too.

One procedure, and it assembles everything this part has covered into one sequence. Name the hypothesis, operationalise both variables, assign at random, specify the control — and then close each confound individually, in turn.

Watch a design get built.

Psych 0101 Designing an Experiment

Naming the hypothesis, operationalising both variables, assigning randomly, specifying the control condition, and closing each confound in turn. The confound step is a list, worked through rather than waved at.

Two entries on a phrase that means something narrow and technical, and gets read as three much larger things. Its actual content is a statement about chance, under a threshold that was set before anyone looked at the data.

What it means.

Psych 0103 Statistical Significance

A result is statistically significant when it is unlikely to have arisen by chance alone, under a threshold set before analysis. The timing of the threshold is part of the definition.

And the misreadings.

Psych 0104 Misreading Statistical Significance

Treating it as proof of importance, of causation, or of a large effect. Three separate misreadings of one narrow claim — and a significant result from a correlational design is still correlational.

Four entries on what stands between a proposal and a study that actually runs. These are not aspirations — the third one is a committee whose approval is required before any data can be collected at all.

The participant's agreement.

Psych 0106 Informed Consent

Agreement to take part after being told the study's risks, procedures, and the right to withdraw. Each of the three has to be told, not assumed.

When purpose is withheld.

Psych 0108 Deception and Debriefing

Deception withholds or misstates a study's purpose during participation; debriefing discloses it fully afterwards. The second is the condition on which the first is permitted.

The gate.

Psych 0110 The Institutional Review Board

A committee reviewing proposed research, which must approve it before data collection begins. Before is the operative word — this is not a review of what happened.

Animal research.

Psych 0112 Ethics in Animal Research

Requiring scientific justification, the minimum necessary numbers, and procedures reducing pain and distress. Three requirements, each independently checkable.

Two entries closing the unit. Running a study again to see whether the result appears again — and then a capstone that takes a published health headline and traces it back through its design to the strongest thing its data actually support.

Running it again.

Psych 0114 Replication

Repeating a study's procedure to see whether the original result appears again. A single result is a candidate finding; a replicated one is closer to a fact.

The capstone. One headline, audited.

Psych 0116 Auditing a Health Headline

Tracing a published headline back through its design to the strongest statement its data actually support — which is very often a weaker statement than the headline made.

Five things you can now do, and Unit One is complete. You can say why correlational data stays correlational — no amount of further reasoning about the same data converts it, because the design decides what the data can support. You can read a correlation coefficient in three steps: sign, strength, and then what the number cannot establish however large it is. You can say what random assignment does that random sampling does not — it builds the groups inside a study and protects the causal claim, while sampling builds the sample and protects the generalisation. You can say what a confound costs: not falsity, but unreadability, because two explanations fit the same data equally well. And you can say what statistical significance establishes — that a result is unlikely by chance under a pre-set threshold — and that it establishes nothing about importance, cause, or size.