Key findings

Key Findings and Effects

The findings and effects of the masterclass, collected unit by unit, each with the study and the year behind it. Every one is explained in its unit, with the data it comes from.

Psychology has a handful of formulas and a great many measured findings. What a psychologist carries in their head is a set of results against which every new claim is judged: how large an effect of 0.5 standard deviations is, how much of a list is forgotten within a day, how many of Milgram's participants went to the last switch, how much testing beats rereading, how far the heritability of intelligence rises with age, and how many patients must be treated for one to benefit. This page lists those findings and effects, in the order of the units, each with the study and the year behind it and, where the result is a formula a student should be able to write down, the formula itself.

Each card states a result the way the Key Finding box of its unit does, and says when it applies or what it does not mean: a benchmark is a convention, a classic figure comes from one sample, a replication may have halved an effect. The page grows as units are published, so a finding you cannot find yet belongs to a unit that is still to come.

Psychology as a Science

The mean, the standard deviation, and the z-score

\[ \bar{x} = \frac{1}{n}\sum_{i=1}^{n} x_i , \qquad s = \sqrt{\frac{1}{n-1}\sum_{i=1}^{n}\left(x_i - \bar{x}\right)^2} , \qquad z = \frac{x - \bar{x}}{s} \] The mean is the average of the \(n\) values, the standard deviation \(s\) is their typical distance from the mean, and the z-score of a value is its distance from the mean in standard deviations. For data that follow a normal distribution, about 68% of the values lie within one standard deviation of the mean, about 95% within two, and about 99.7% within three (the 68-95-99.7 rule). For skewed data the mean is pulled toward the tail and the median is the better summary of a typical value.

Pearson’s correlation coefficient

\[ r = \frac{1}{n-1}\sum_{i=1}^{n} z_{x,i}\, z_{y,i} \] The average product of the z-scores of two variables measured in the same \(n\) people. It runs from \(-1\) (a perfect falling line) through 0 (no linear relation) to \(+1\) (a perfect rising line), and \(r^2\) is the share of the variance of one variable that a straight line through the other accounts for. A correlation says how strongly two variables go together and nothing about which one causes the other.

Cohen’s d

\[ d = \frac{\bar{x}_1 - \bar{x}_2}{s_{\mathrm{pooled}}} , \qquad s_{\mathrm{pooled}} = \sqrt{\frac{(n_1 - 1)\,s_1^2 + (n_2 - 1)\,s_2^2}{n_1 + n_2 - 2}} \] The difference between two group means in standard deviations, where \(s_{\mathrm{pooled}}\) combines the two groups’ standard deviations with weights given by their sizes. Cohen suggested calling \(0.2\) small, \(0.5\) medium, and \(0.8\) large, and these benchmarks are used everywhere, with the caution that the context decides: a \(d\) of \(0.2\) for a cheap treatment given to millions of patients is enormous, and a \(d\) of \(0.8\) in a laboratory task may matter to nobody.

The p-value

The \(p\)-value of a result is the probability of obtaining data at least as extreme as the data observed if the null hypothesis were true. By a convention that goes back to Ronald Fisher (1890-1962), a result with \(p < 0.05\) is called statistically significant. The threshold is a convention and nothing more; nature does not change at \(0.05\), and a result with \(p = 0.04\) and one with \(p = 0.06\) are almost the same evidence.

The Reproducibility Project: Psychology

Of 100 studies published in three leading psychology journals in 2008, 97 had reported a significant result; when they were repeated with larger samples, 36% of the replications (35 of 97) were significant in the same direction, and the replication effect sizes were on average about half the size of the original ones (Open Science Collaboration 2015). The figures describe those journals and that year, and the two subfields differed: cognitive findings replicated about twice as often as social ones.

Neurons, Brain, and Behaviour

The Nernst equation

\[ E_{\mathrm{ion}} = \frac{61.5\unit{mV}}{z}\,\log_{10}\!\left(\frac{c_{\mathrm{out}}}{c_{\mathrm{in}}}\right) \qquad \text{at } 37\,^{\circ}\mathrm{C} \] The equilibrium potential \(E_{\mathrm{ion}}\) of an ion is the voltage across the membrane (inside relative to outside) at which diffusion down the concentration gradient and the electrical force exactly cancel. \(c_{\mathrm{out}}\) and \(c_{\mathrm{in}}\) are the concentrations outside and inside the cell, and \(z\) is the charge of the ion, \(+1\) for K\(^+\) and Na\(^+\), \(-1\) for Cl\(^-\), and \(+2\) for Ca\(^{2+}\). The constant \(61.5\unit{mV}\) holds at body temperature. The formula describes one ion species on its own; the real membrane potential is a compromise between the ions the membrane lets through.

The action potential

A resting neuron sits at about \(-70\unit{mV}\). When its membrane is pushed up to a threshold near \(-55\unit{mV}\), an action potential follows: the potential shoots up to a peak near \(+40\unit{mV}\) and falls back within about \(1\unit{ms}\), briefly undershooting the resting value. The spike travels along the axon at speeds from under \(1\unit{m\,s^{-1}}\) in thin unmyelinated fibres to about \(120\unit{m\,s^{-1}}\) in the thickest myelinated ones.

The brain in numbers

The adult human brain weighs about \(1.5\unit{kg}\), as much as a litre and a half of water and about 2% of body mass, and it uses about 20% of the energy of the resting body. It contains about 86 billion neurons and about 85 billion non-neuronal cells, so glia are at most about half of all its cells, not nine tenths (Azevedo et al. 2009). The cerebral cortex makes up 82% of the brain’s mass but holds only 19% of its neurons, about 16 billion; the cerebellum makes up 10% of the mass and holds 69 billion neurons, four of every five in the brain.

Sensation and Perception

Sensitivity \(d'\) and the criterion

\[ d' = z(H) - z(F) , \qquad c = -\tfrac{1}{2}\,\bigl[\,z(H) + z(F)\,\bigr] \] \(H\) is the hit rate, \(F\) the false alarm rate, and \(z(\cdot)\) turns a proportion into the \(z\)-score of the normal curve of Unit 1, that is, into the number of standard deviations above the mean below which that proportion of the curve lies, so that \(z(0.5) = 0\), \(z(0.84) = 1\) (84% of the curve lies below one standard deviation above the mean), and \(z(0.16) = -1\). The sensitivity \(d'\) is the distance between the means of the two distributions in standard deviations and does not change when the observer becomes more or less willing to say “yes”. The criterion \(c\) is zero for an unbiased observer, positive for a cautious one, and negative for a liberal one. The formulas assume that both distributions are normal with the same spread.

The eye in numbers

A human retina holds about 90 to 120 million rods and about 4.5 to 6.5 million cones, and about one million axons leave each eye in the optic nerve. Only the fovea sees sharply, and sharpest in its centre, the central one to two degrees of the visual field; the eyes make about three to four saccades per second to compensate. The blind spot of each eye is about 5 to 6 degrees wide and 7 to 8 degrees high, about 15 degrees to the outer side of the point of gaze, and is never seen because the other eye and the brain fill it in. A rod responds to a single photon (Hecht, Shlaer, and Pirenne 1942).

Perception is inference

What a person perceives is the visual system’s best estimate of the scene that produced the image on the retina, computed from the image together with assumptions learned about the world, rather than a copy of the image. The two best pieces of evidence are the illusions of size, in which cues of distance drawn on paper make equal lines look unequal and the Ames room makes equal people look unequal, and the dress photograph of 2015, in which the same pixels were seen as blue and black by 57% of 1,401 people and as white and gold by 30%, depending on the illumination each visual system assumed (Lafer-Sousa, Hermann, and Conway 2015).

Learning and Conditioning

Extinction is new learning, and the old learning survives

Repeating a conditioned stimulus without its unconditioned stimulus makes the conditioned response decline, but the response returns after a rest (spontaneous recovery), in a new context (renewal), and after a few presentations of the unconditioned stimulus alone (reinstatement). In Pavlov’s 1927 demonstration the metronome fell from 10 drops to 3 over seven unreinforced trials and gave 6 drops again after 23 minutes. Extinction does not delete the learned association; it lays a second, inhibitory one over it.

Blocking (Kamin 1969)

When a stimulus is added to one that already predicts the unconditioned stimulus, and the compound is reinforced, the added stimulus acquires little or no conditioned response, although it has been paired with the unconditioned stimulus as often as in a control group trained on the compound alone. With the contingency experiments of Rescorla (1968), blocking shows that conditioning is learning to predict: a predicted outcome teaches nothing new.

Taste aversion learning (Garcia and Koelling 1966)

A taste followed by illness becomes aversive after one pairing, even when the illness begins hours later, whereas lights and sounds followed by the same illness are not learned, and the reverse holds for pain. Conditioning is constrained by what each species is built to connect.

Reinforcement and punishment

Reinforcement makes a behaviour more frequent, either by adding something (positive reinforcement) or by removing something (negative reinforcement, such as taking an aspirin that ends a headache). Punishment makes a behaviour less frequent, either by adding something (positive punishment) or by removing something (negative punishment, such as a fine). Negative reinforcement strengthens the behaviour that escapes from an unpleasant event, and it is what keeps avoidance going in a phobia, because every escape is reinforced by the relief it brings.

The partial reinforcement extinction effect (Humphreys 1939)

A behaviour that has been reinforced only on some occasions is more resistant to extinction than one that has been reinforced every time, and a variable ratio schedule produces the most persistent behaviour of all. The effect explains why a child whose tantrums are given in to only occasionally goes on having them the longest, and why a gambler keeps playing a machine that pays out at unpredictable intervals.

Memory and Forgetting

The capacity of short-term and working memory

An adult holds about four chunks in the focus of attention when rehearsal and grouping are prevented (Cowan 2001), and repeats back about seven items, plus or minus two, in a span test that allows them (Miller 1956). Without rehearsal the content is largely lost within about 20 seconds (Peterson and Peterson 1959). A chunk is a unit of meaning, so the same four slots hold four letters or four well-known words, and practice can pack more into each chunk but does not add slots.

The dissociation of explicit and procedural memory in H.M.

After the removal of the medial temporal lobes on both sides in 1953, patient H.M. could no longer form lasting memories of events and facts (anterograde amnesia) and had lost memories of the years just before the operation, while his intelligence, language, short-term memory, and skill learning were spared (Scoville and Milner 1957; Milner 1962). Explicit memory and procedural memory therefore depend on different parts of the brain.

The serial position effect

In free recall of a list, the first few items (the primacy effect) and the last few (the recency effect) are recalled better than the middle. A 30-second delay filled with counting removes the recency effect and leaves the primacy effect (Glanzer and Cunitz 1966), which shows that the last items are recalled from short-term memory and the first from long-term memory. The effect was mapped by Murdock (1962) with lists of 10 to 40 words.

The forgetting curve

Forgetting is fast at first and slow later. In Ebbinghaus’s (1885) data the savings on relearning a list fell to 58% after 20 minutes, 44% after an hour, 34% after a day, and 21% after 31 days, so most of a month’s loss occurs within the first day, and the curve is close to a straight line on a logarithmic time axis. The same shape has been found for prose, faces, and school knowledge, with slower rates for meaningful material.

The testing effect

Recalling material from memory strengthens it more than restudying it for the same time, and the advantage grows with the delay before the final test. In Roediger and Karpicke’s (2006) second experiment, students who read a passage once and recalled it three times remembered 61% of it a week later, against 40% for students who read it four times, although the readers had predicted the opposite. Testing works without feedback; with feedback it works better still.

The misinformation effect

Information that reaches a person after an event can become part of the memory of the event. In Loftus and Palmer (1974), the verb smashed in a question about speed raised the estimate from 31.8 (contacted) to 40.8 miles per hour, and a week later 32% of the students who had read smashed remembered broken glass that was not in the film, against 14% of those who had read hit.

Thinking, Language, and Decisions

The Wason selection task

Asked which of four cards to turn to test a rule of the form “if P then Q”, about 10% of people choose the two cards that can falsify it, P and not Q. Most choose P alone or P and Q, which can only confirm the rule. When the same logical structure is given a familiar social content, such as a drinking-age rule, about three quarters of people choose correctly (Griggs and Cox 1982).

Anchoring

In the study of Tversky and Kahneman (1974), a wheel of fortune stopped at 10 for one group and at 65 for another before each group estimated the percentage of African countries among the members of the United Nations. Each participant first said whether the true percentage was higher or lower than that number and then estimated it by moving up or down from it. The median estimates were 25% and 45%: both groups moved in the right direction but stopped early, so a number they had watched being produced by chance moved their estimates by 20 percentage points, and paying for accuracy did not reduce the effect. Anchoring was replicated in all 36 samples of the Many Labs project (Klein et al. 2014), and anchors have since been shown to move the valuations of houses by estate agents and the sentences handed down by judges in experiments with a rolled die.

Bayes’ theorem for a diagnostic test

\[ P(D \mid +) = \frac{P(+ \mid D)\,P(D)}{P(+ \mid D)\,P(D) + P(+ \mid \text{not } D)\,P(\text{not } D)} \] The probability of the disease after a positive test equals the share of all positive tests that come from sick people. In natural frequencies: of 1,000 people, count how many are sick and test positive, count how many are healthy and test positive, and divide the first number by the sum of the two.

The Stroop effect

Naming the ink colour of a colour word printed in a conflicting colour takes longer than naming the colour of a plain patch: in Stroop’s 1935 experiment with 100 participants the difference was 47 seconds per 100 items, about \(470\unit{ms}\) per word, while the ink colour barely slowed the reading of the words. Reading is automatic, and attention can only manage the conflict, not prevent it.

Social Influence and Group Behaviour

The fundamental attribution error

Observers explain other people’s behaviour by their dispositions and underweight the situation. In Jones and Harris’s (1967) experiment, readers who knew that a writer had been assigned the side of an essay about Castro still rated the writer’s attitude 21 points apart on a 60-point scale depending on the side, half the gap they attributed to writers who had chosen freely. The effect is one of the most replicated in social psychology, and it is larger in individualist cultures, which stress the independent person, as in North America and Western Europe, than in collectivist ones, which stress the group, as in much of East Asia.

Asch’s conformity effect

When a unanimous majority gives an obviously wrong answer to an easy perceptual question, people follow it on about a third of the trials: 32% of critical answers in Asch’s 1951 study with 50 participants, and 36.8% in the 1956 report with 123, of whom about a quarter never yielded. A majority of three produces almost the full effect, and one ally who answers correctly cuts the error rate to about a quarter of its value.

Milgram’s obedience finding

In the baseline study of Milgram (1963), 40 men were instructed by an experimenter to give what they believed were increasingly painful electric shocks to a protesting learner. All 40 went to at least 300 volts, and 26 of them, which is 65%, went to the last switch, marked 450 volts. The mean maximum shock was 405 volts, and the obedience rate changed with the situation far more than with the person.

The bystander effect

The more people witness an emergency, the less likely each of them is to help and the longer they take. In Darley and Latané (1968), 85% of participants who believed they were the only witness reported a seizure before the victim fell silent, against 62% with one other presumed witness and 31% with four. Two mechanisms carry the effect: pluralistic ignorance, in which calm bystanders define the event as harmless, and diffusion of responsibility, in which each person’s share of the duty to act shrinks with the number present.

The minimal group effect

Dividing people into groups on a trivial or random basis is enough to produce favouritism towards the in-group. In Tajfel et al. (1971), schoolboys sorted as “overestimators” or “underestimators” of dots, who never met their groups and gained nothing themselves, allocated more money to members of their own group, and in a second experiment chose to maximize the difference between the groups even at a cost to their own.

The Big Questions of Psychology

Heritability and the Falconer formula

\[ h^2 = 2\,(r_{\mathrm{MZ}} - r_{\mathrm{DZ}}), \qquad c^2 = 2\,r_{\mathrm{DZ}} - r_{\mathrm{MZ}}, \qquad e^2 = 1 - r_{\mathrm{MZ}} \] Heritability \(h^2\) is the share of the variance of a trait in a population that is associated with genetic differences; \(c^2\) is the share due to the environment that twins share, and \(e^2\) the share due to everything the twins do not share, including measurement error. \(r_{\mathrm{MZ}}\) and \(r_{\mathrm{DZ}}\) are the correlations between identical and between fraternal twins. The formula assumes that the genetic effects add up, that both kinds of twins share their environment equally, and that genes and environment are not correlated; a modern twin study, which fits a statistical model to the data instead of using the formula, can test these assumptions rather than assume them.

Libet’s timing

In Libet et al. (1983) the readiness potential began about 550 ms before a spontaneous voluntary movement, and the participants reported the urge to move about 200 ms before it. Brain activity preceding the movement therefore began about 350 ms before the participants were aware of deciding. The finding is reliable and has been repeated many times; what it means is disputed.

The replication record and the reform

When 100 studies from three leading journals were repeated, 36% of the 97 significant originals replicated with a significant result and the effects were about half as large (Open Science Collaboration 2015). The causes have names, p-hacking, HARKing, publication bias, and the garden of forking paths, and the remedies address them: preregistration, Registered Reports, in which 44% of first hypotheses are confirmed against 96% in the ordinary literature (Scheel et al. 2021), multi-site replication, larger samples, and open data.

Emotion and Motivation

The two-factor theory and the Schachter and Singer experiment

An emotion is the product of two factors, a state of physiological arousal and a cognition that labels it. Schachter and Singer (1962) offered 185 male students an injection under the cover story of a vitamin test, which all but one accepted, gave them adrenaline or a saline placebo, told some the true side effects, some nothing, and some false side effects, and then left each with an actor who behaved either euphorically or angrily. Participants who had the arousal but no explanation for it took over the actor’s mood by self-report and by behaviour; participants who had been told what to expect did not.

Universality and its limits

Among the Fore, participants who had never seen Western faces chose the intended photograph for happiness, anger, sadness, disgust, and surprise at rates of 65% to 100% against a chance level of 33%, and did not separate fear from surprise (Ekman and Friesen 1971). When Himba participants sorted the same kind of faces freely, without words, their sorts matched the posed categories 49.6% of the time against 67.1% for Americans, and only smiles, wide-eyed faces, and neutral faces formed consistent piles (Gendron et al. 2014). The method decides how universal the faces look: forced choice with words gives high agreement, and free sorting gives agreement only for a few expressions.

The Whitehall gradient

Among 17,530 London civil servants followed for seven and a half years, deaths from coronary heart disease rose step by step down the four employment grades, and the men in the lowest grade, the messengers, died of it at 3.6 times the rate of the administrators at the top (Marmot et al. 1978). The lower grades were shorter, heavier for their height, smoked more, and had higher blood pressure and blood glucose, yet when all of these and cholesterol were allowed for, the association stayed strong: in the ten-year follow-up the classic risk factors accounted for no more than about 40% of the grade difference, and Marmot has since put their share at about a third (Marmot, Shipley, and Rose 1984). The gradient runs through the whole hierarchy, not only between the poor and the rest, and a lower position in it predicts earlier death even among people who are neither poor nor in danger.

Wanting is not liking

The motivation to obtain a reward (wanting, carried mainly by dopamine) and the pleasure of consuming it (liking, carried by opioid signals in small hotspots of the nucleus accumbens and ventral pallidum) are separate processes. Rats depleted of up to 99% of their striatal dopamine stop seeking food but still show normal pleasure reactions to sweet tastes (Berridge and Robinson 1998). Addiction is the extreme case: with repeated drug use, wanting is sensitized and grows while liking stays the same or falls. How dopamine neurons compute what to want lies beyond this unit and is taken up in a later part of this masterclass.

The overjustification effect

A reward that is expected and given for doing an activity reduces the intrinsic motivation for that activity once the reward is withdrawn. In Lepper, Greene, and Nisbett (1973), nursery children who drew for a promised certificate later spent 8.59% of their free time drawing, against 16.73% for children who had drawn for nothing and 18.09% for children rewarded unexpectedly. Across 128 experiments, expected tangible rewards reduced free-choice engagement by an effect size of \(d = -0.40\) (in units of the standard deviation, as in Unit 1) when given for taking part and \(-0.28\) when given for performing well, whereas praise raised it by \(d = 0.33\) (Deci, Koestner, and Ryan 1999).

Development Across the Lifespan

The strange situation patterns

In Ainsworth’s strange situation, infants of 12 to 18 months are classified from their behaviour when the mother returns. About two thirds are secure, one in five avoidant, and one in seven resistant (65%, 21%, and 14% across 1,990 infants in eight countries; van IJzendoorn and Kroonenberg 1988), with a fourth, disorganized pattern scored separately. The pattern reflects the history of the relationship, and samples within one country differ more than countries do.

Object permanence by looking against object permanence by searching

Infants look longer at an event in which a hidden object fails to block a moving screen than at one in which it blocks it. They do so at 4 and a half months, and some of them at 3 and a half (Baillargeon 1987; mean looking times of \(29.2\) against \(17.7\) seconds at 4 and a half months), whereas they first search for a fully hidden toy at about 8 to 9 months, and Piaget placed complete object permanence at 18 to 24 months. Piaget’s sequence of stages has held; his ages were set by tasks that demanded action and language, and looking-time tasks show the knowledge months earlier.

The false-belief shift

Children begin to answer the false-belief question correctly at about 4 years of age, moving from below-chance to above-chance performance across the preschool years in 178 studies (Wellman, Cross, and Watson 2001). In the Sally-Anne task of Baron-Cohen, Leslie, and Frith (1985), 23 of 27 typically developing 4-year-olds and 12 of 14 children with Down syndrome answered correctly, against 4 of 20 autistic children.

The adolescent brain and the peer effect

Cortical grey matter falls and white matter rises through adolescence into the twenties, latest in the frontal and temporal lobes, while the subcortical reward system becomes more responsive at puberty. In Gardner and Steinberg’s (2005) study of 306 people, the presence of two friends about doubled risk taking in a driving game for 13- to 16-year-olds, raised it by about half for 18- to 22-year-olds, and did not change it for adults; alone, the three age groups did not differ.

The fluid-crystallized crossover

Fluid abilities (reasoning, memory for new material, spatial ability, and speed) are highest in the twenties and decline nearly linearly through adult life, by one and a half to two standard deviations between 20 and 90 in cross-sectional data. Crystallized abilities (vocabulary and general knowledge) rise until about 60 and decline only in late old age. The two curves cross near the age of 50 (Salthouse 2010; Hartshorne and Germine 2015), and longitudinal data show the fluid decline starting later and running more gently than the cross-sectional curves suggest.

Personality and Intelligence

The comparative validity of personality

In prospective longitudinal studies, the Big Five predict mortality, divorce, and occupational attainment about as well as socioeconomic status and intelligence do. Conscientiousness predicts a longer life (\(r = 0.09\)) at least as well as intelligence (\(0.06\)) and better than family socioeconomic status (\(0.02\)); neuroticism and low agreeableness predict divorce (\(r = 0.17\) and \(0.18\)) better than socioeconomic status (\(0.05\)). All of these effects are small by the conventions of Unit 1, and all of them are of the same order (Roberts et al. 2007).

The IQ scale and the Flynn effect

Wechsler IQ scores are placed on a normal distribution with mean 100 and standard deviation 15, so \(\mathrm{IQ} = 100 + 15z\), and the share of people above a given score follows from the normal curve: about 16% score above 115, about 2.3% above 130, and about 0.1% above 145. The norms are set by the sample of the year in which the test was standardized. Raw performance on the same test has risen by about 3 IQ points per decade through most of the twentieth century in many countries, which is the Flynn effect, and the rise has slowed, stopped, or reversed in some countries since the 1990s.

What an IQ score predicts

Intelligence test scores correlate about 0.5 with school grades (0.44 observed, 0.54 corrected, Roth et al. 2015) and about 0.5 with performance in jobs of medium complexity (Schmidt and Hunter 1998), more in complex jobs and less in simple ones, and a one standard deviation advantage in youth goes with a 24% lower risk of death in the following decades (Calvin et al. 2011). Because the square of a correlation is the share of the variance accounted for (Unit 1), a correlation of 0.5 accounts for a quarter of the variance in grades or performance and leaves three quarters to other things, among them the conscientiousness of Section 2.

The Falconer formula

\[ h^2 = 2\,(r_{\mathrm{MZ}} - r_{\mathrm{DZ}}), \qquad c^2 = 2\,r_{\mathrm{DZ}} - r_{\mathrm{MZ}}, \qquad e^2 = 1 - r_{\mathrm{MZ}} \] Here \(r_{\mathrm{MZ}}\) and \(r_{\mathrm{DZ}}\) are the correlations of a trait between identical (monozygotic) and fraternal (dizygotic) twins reared together, \(h^2\) is the heritability, \(c^2\) the share of the variance due to the environment the twins share, and \(e^2\) the share due to the environment they do not share, together with measurement error. The formula assumes that the shared environment is equally shared by both kinds of twins and that genetic effects add up, so that each gene’s effect adds to the others’ instead of depending on them. The heritability of intelligence estimated this way rises from about 0.2 in infancy to about 0.6 to 0.8 in adulthood (Plomin et al. 2016).

The gender similarities hypothesis

Across 46 meta-analyses, 78% of the psychological differences between women and men had a Cohen’s \(d\) of 0.35 or less, and 30% had a \(d\) of 0.10 or less (Hyde 2005). Large differences are confined to a few domains, among them throwing and physical aggression. A \(d\) of 0.35 means that the two distributions overlap by about 86% (by the overlap of two normal curves shown in Unit 1), so that most women and most men lie in the range the sexes share.

Psychological Disorders and Their Treatment

What a psychological disorder is

In the DSM-5-TR and the ICD-11, a mental disorder is a clinically significant disturbance of cognition, emotion regulation, or behaviour that reflects a dysfunction in the processes behind mental functioning and is usually accompanied by distress or by impairment in daily life. Expectable responses to loss or stress, and disagreements between a person and their society, are excluded. The definition follows Wakefield’s (1992) harmful dysfunction account: a failure of a mental mechanism (a matter of fact) that causes harm by the standards of the person’s culture (a matter of value).

How common psychological disorders are

In the National Comorbidity Survey Replication (Kessler et al. 2005), 46.4% of US adults had met the criteria for at least one disorder in their lifetime and 26.2% in the past year; half of all lifetime cases had begun by age 14. The Global Burden of Disease study estimated 280 million people with a depressive disorder and 301 million with an anxiety disorder worldwide in 2019. Across 21 countries, 16.5% of people with major depression in the past year received minimally adequate treatment (Thornicroft et al. 2017).

The heritability of schizophrenia

Twin studies put the heritability of schizophrenia at about 0.8 (0.79 in the Danish register study of Hilker et al. 2018; 0.81 in the meta-analysis of Sullivan et al. 2003). The concordance is about 33% in identical twins and 7% in fraternal twins, so most identical co-twins of a patient never develop the disorder. Heritability describes the share of the variation in a population associated with genetic differences, in that population and that environment; it says nothing about any one person and does not mean that the disorder is 80% genetic in anyone.

The number needed to treat

\[ \mathrm{NNT} = \frac{1}{p_{\mathrm{treatment}} - p_{\mathrm{control}}} \] Here \(p_{\mathrm{treatment}}\) and \(p_{\mathrm{control}}\) are the shares of patients who respond in the two groups of a randomized trial, and their difference is the absolute benefit of the treatment. The result is rounded up to a whole number of patients. A relative measure such as an odds ratio, which compares the odds \(p/(1-p)\) of responding in the two groups, says nothing about the absolute benefit until the control rate is known.

What the treatments for depression achieve

Against care as usual, the main psychotherapies for adult depression have standardized mean differences of about 0.5 to 0.8 and do not differ reliably from each other, with supportive counselling weaker at 0.32 (Cuijpers et al. 2021, 331 trials). Against placebo, the 21 antidepressants in Cipriani et al. (2018, 522 trials) have odds ratios of 1.37 to 2.13 for response and a standardized mean difference of 0.30, which with a placebo response of 30% means numbers needed to treat of about 6 to 15. Combining a psychotherapy with a drug produces response about 1.25 times as often as either alone (Cuijpers et al. 2020).

Brain Imaging and Stimulation

The timing of the BOLD response

After a brief burst of neural activity the BOLD signal of the region begins to rise after 1 to 2 s, peaks 4 to 6 s after the activity, returns to its baseline after about 12 s, and dips below it for a further 15 to 20 s. The response to a longer stimulus is approximately the sum of such responses, so a block of stimulation lasting 20 s produces a plateau that begins and ends several seconds late.

Expected false positives

\[ E = N\,\alpha \] When \(N\) independent tests are made at a significance level \(\alpha\) and no effect exists, the expected number of tests that pass is \(N\alpha\), and the probability of at least one false positive is \(1 - (1-\alpha)^N\), since \((1-\alpha)^N\) is the probability that every one of the \(N\) tests stays quiet; this is close to 1 for any whole-brain analysis. The Bonferroni correction tests each voxel at \(\alpha/N\) and keeps the probability of any false positive at \(\alpha\); because neighbouring voxels are correlated, it is stricter than necessary.

Cluster inference and the finding of Eklund, Nichols, and Knutsson (2016)

Using resting-state scans of 499 healthy people, in which no task-related activity existed, the authors ran about three million random group analyses with the three most used software packages. Voxel-wise familywise correction was valid, even conservative. Cluster inference with a cluster-defining threshold of \(p = 0.01\) produced at least one false cluster in up to \(70\%\) of analyses instead of the nominal \(5\%\); a threshold of \(p = 0.001\) came much closer to \(5\%\), and a permutation test, which makes no assumption about the shape of the noise, held the rate at \(5\%\). The cause was that the noise in real fMRI data stays correlated over longer distances than the assumed bell-shaped, or Gaussian, pattern allows, a property statisticians call heavy tails.

The NARPS result (Botvinik-Nezer et al. 2020)

No two of the 70 teams used the same analysis pipeline. For one hypothesis, \(84.3\%\) of the teams reported a significant effect; for three hypotheses, \(5.7\%\) did; and for the remaining five hypotheses the teams split, with between \(21.4\%\) and \(37.1\%\) reporting a significant effect. Averaged over the nine hypotheses, one team in five reported the opposite of the majority. Combining the unthresholded maps of all teams in a meta-analysis gave a consistent result, and researchers who bet on the outcome in a prediction market overestimated how many teams would find significant effects.

Memory in the Brain

Long-term potentiation

Long-term potentiation (LTP) is a lasting increase in the strength of a synapse after a brief burst of high-frequency activity, first reported in the rabbit dentate gyrus by Bliss and Lømo (1973), where it lasted from 30 minutes to 10 hours. It has three properties that match Hebb’s rule.

  1. Cooperativity: a weak input to one or a few fibres does not induce it; enough fibres must be active together to depolarize the postsynaptic cell.
  2. Associativity: a weak input is potentiated if it is active at the same time as a strong input to the same cell.
  3. Input specificity: only the synapses that were active during the burst are strengthened; other synapses on the same cell are unchanged.

Place cells and grid cells

Place cells in the hippocampus (O’Keefe and Dostrovsky 1971) fire in one region of an environment each, and grid cells in the medial entorhinal cortex (Hafting et al. 2005) fire on a hexagonal lattice of regions whose spacing grows from the dorsal to the ventral end of the structure. Together with head direction cells and border cells they form the brain’s positioning system, for which John O’Keefe, May-Britt Moser, and Edvard Moser received the Nobel Prize in Physiology or Medicine in 2014.

The standard model of systems consolidation and the retrograde gradient

A new declarative memory depends on the hippocampus, which binds its parts in the neocortex; with time and repeated reactivation the cortical parts become linked directly, and the memory no longer needs the hippocampus. Hippocampal damage therefore causes a retrograde amnesia that is worst for the years just before the damage and absent for remote years, the gradient that Ribot described in 1881. In rats the gradient is measured in weeks: a hippocampal lesion one day after fear conditioning abolished the fear of the training context (the chamber in which the shocks had been given), whereas lesions 7, 14, or 28 days later spared more of it the longer the delay, and the 28-day group was close to normal (Kim and Fanselow 1992).

Replay and targeted memory reactivation

Place cells that fired together during behaviour fire together again in the slow-wave sleep that follows (Wilson and McNaughton 1994), and a memory can be reactivated on purpose by presenting, during slow-wave sleep, a cue that was present during learning. In Rasch et al. (2007) the smell of roses during slow-wave sleep raised overnight recall of card positions from 85.8% to 97.2%; the cue did nothing during REM sleep or wakefulness. Across 91 later experiments the average benefit is \(g = 0.29\) (Hu et al. 2020).

The engram tagged and switched

The neurons of the dentate gyrus that are active during an experience can be labelled with a light-sensitive channel, and firing them again with light is enough to produce the behaviour that the memory produces (Liu et al. 2012). Reactivating the cells of a safe place while a mouse is shocked elsewhere gives the mouse a fear of the safe place: about 25% freezing there against 3% before (Ramirez et al. 2013). A memory is held by a specific, sparse set of cells, and their activity is sufficient for recall.

The Predictive Brain

Bayes’ theorem for perception

\[ P(\mathrm{scene} \mid \mathrm{image}) = \frac{P(\mathrm{image} \mid \mathrm{scene}) \, P(\mathrm{scene})}{P(\mathrm{image})} \] The vertical bar is read “given”. The posterior probability of a scene given the image is proportional to the likelihood of the image under that scene multiplied by the prior probability of the scene. The denominator is the same for every candidate scene and only makes the posteriors add up to 1, so the product of likelihood and prior decides. When the input is sharp and reliable the likelihood dominates; when it is weak or ambiguous the prior takes over.

The mismatch negativity

A rare deviant tone in a sequence of identical standards produces a negative wave at frontal and central electrodes, measured as the deviant-minus-standard difference, that begins about 150 ms and peaks about 190 ms after the tone. In the ERP CORE dataset of Kappenman et al. (2021), 39 young adults who ignored the tones while watching a silent video showed a mean peak of \(-3.46\unit{\mu V}\) (microvolts, millionths of a volt) at 188 ms, with an onset at 147 ms and an effect size of \(d_z = 2.02\) against zero (the mean peak divided by its standard deviation across participants). The wave is present without attention, which makes it the standard evidence that the auditory cortex predicts the next sound and signals a violation of that prediction. Näätänen, Gaillard, and Mäntysalo (1978) reported it first.

The dopamine reward prediction error (Schultz, Dayan, and Montague 1997)

Dopamine neurons of the midbrain respond in three ways. An unpredicted reward produces a burst of firing. A reward that is fully predicted by a cue produces no response, and the burst appears instead at the cue, the earliest event that predicts the reward. A predicted reward that is omitted produces a pause in firing at the moment the reward was due. The three responses are what a single quantity would do, the reward prediction error, the value of what was received minus the value of what was predicted. The neurons do not signal reward or pleasure; they signal how much better or worse than expected the moment turned out.

The size of the placebo effect in pain

Measured against no treatment, the placebo effect depends on the expectation that was created. In clinical trials in which the placebo was only a control, its mean effect on pain was \(d = 0.15\); in experiments that induced an expectation of relief by suggestion or conditioning it was \(d = 0.95\) (Vase, Riley, and Price 2002, 23 and 14 studies). Across 202 trials with a no-treatment group, the Cochrane review of Hróbjartsson and Gøtzsche (2010) found a standardized mean difference of \(-0.28\) for pain (negative because the placebo groups reported less pain). The effect is real, is blocked by naloxone when it is opioid-mediated, and is large only where a prediction of relief has been formed.

Conditioned hallucinations and strong priors (Powers, Mathys, and Corlett 2017)

After a visual pattern has been paired with a faint tone, people report the tone on trials on which none is played. All four groups, 59 participants in all, showed the effect; the 30 who heard voices in daily life, with or without a psychiatric diagnosis, showed it far more often than the 29 who did not, were more confident when they did, and showed activity in the auditory cortex while doing so. A model fitted to the responses attributed the difference to a stronger weighting of the prior expectation, not to poorer hearing. The finding supports the account of hallucination as a prior that overrides the senses.

Consciousness and the Brain

Sensitivity and criterion applied to awareness

\[ d' = z(H) - z(F) , \qquad c = -\tfrac{1}{2}\,\bigl[\,z(H) + z(F)\,\bigr] \] When the report is “seen” or “not seen”, \(H\) is the share of stimulus trials reported seen, \(F\) the share of blank trials reported seen, and \(z(\cdot)\) the \(z\)-score of the normal curve, that is, the point on the standard normal curve, measured in standard deviations from the mean, below which the given share of the area lies. A stimulus is below the objective threshold when \(d' = 0\) for detecting it in a forced choice, and below the subjective threshold when the person reports not seeing it although \(d' > 0\). Only the first shows that nothing got in; the second may be perception without awareness or only a cautious criterion, and the two cannot be told apart without measuring both.

Awareness detected without behaviour

In Owen et al. (2006), one patient who met every clinical criterion of the vegetative state produced sustained, task-specific brain activity when asked to imagine playing tennis and to imagine walking through her home, indistinguishable from that of 12 healthy volunteers. In Monti et al. (2010), 5 of 54 patients with disorders of consciousness (23 vegetative, 31 minimally conscious) could modulate their brain activity in this way, 4 of them with a diagnosis of vegetative state on admission, and one of these answered five of six yes-or-no questions about his own life correctly by imagining tennis for “yes” and his home for “no”. A negative result does not show the absence of awareness: 49 of the 54 patients, among them most of the minimally conscious ones, produced no detectable response.

The ignition signature

When the same stimulus is compared on seen and unseen trials, the first 200 ms of the brain’s response are alike. Seen trials then show a late, widespread response that does not grow gradually with the stimulus (a non-linear response), beginning at about 270 ms and most visible as the P3 wave and as sustained activity in prefrontal and parietal cortex, which is absent or decaying on unseen trials (Sergent, Baillet, and Dehaene 2005; Del Cul, Baillet, and Dehaene 2007). The response is all or none across the threshold, which the global neuronal workspace theory predicts and calls ignition. Whether the whole of this late response is the correlate of experience or partly of its report is disputed.

The 2025 adversarial test of two theories

In the Cogitate study (2025), the category of a seen picture could be decoded from posterior cortex for as long as the picture was on the screen and from the prefrontal cortex only briefly and incompletely, and the orientation of a seen face could be decoded from posterior cortex but not from the prefrontal cortex. Posterior activity tracked the duration of the picture in only 25 of 657 posterior recording sites, and the sustained synchrony between posterior areas that integrated information theory had predicted was not found. The prefrontal cortex responded at the onset of the picture, but none of the 655 prefrontal recording sites showed the response at its offset that the workspace theory had predicted. Each theory passed some of its tests and failed others it had agreed in advance would count against it: the workspace theory the ignition at offset and the prefrontal decoding of orientation, integrated information theory the sustained posterior synchrony.

The readiness potential after Schurger

In Libet et al. (1983) the readiness potential began about 550 ms before a spontaneous voluntary movement and the participants reported the urge to move about 200 ms before it. Schurger, Sitt, and Dehaene (2012) showed that a slow ramp of exactly this shape is produced when random fluctuations of motor activity are averaged backwards from the moment they happen to cross a threshold. The reason is that the stretches just before a crossing are, on average, stretches in which the activity happened to be rising, so the average of many of them rises smoothly even though no single one contains a decision. Their model also predicted correctly when people would respond fastest to an unexpected signal. The readiness potential is therefore not evidence that the brain decides before the person is aware of deciding; it is what the average of many undecided moments looks like, and a prepared movement can still be cancelled up to about 200 ms before it would begin (Schultze-Kraft et al. 2016). The timing experiments show that a voluntary act is prepared gradually by processes the person does not experience, not that the experience of deciding is a bystander.