Methods — Edexcel A-Level Psychology

Design starts with the aim, after which each item is written to measure one thing at a time in plain language, avoiding leading wording, double negatives, jargon and items that ask two things at once. A pilot on a handful of people shows which items are misread, and standardised instructions with a fixed order keep the procedure the same for everyone. The researcher is part of the measurement, because age, sex, ethnicity, accent, dress, tone and even a nod can shift what a participant is willing to say, and expectancy shows up in how an ambiguous answer gets coded. Adorno's team is the cautionary example here, since interviewers already knew the scale scores before they interpreted the interviews. Controls include a written schedule, trained interviewers kept blind to the aim, recording with a second independent coder, and anonymous completion for sensitive topics.

The three interview formats trade control against depth. A structured interview uses a fixed schedule, giving comparable data that is easy to replicate, but cannot follow unexpected leads. A semi-structured format keeps core questions while allowing the interviewer to probe, favoured in clinical settings. An unstructured format uses only a topic list, yielding rich material analysed thematically, though reliability suffers. Question types share this trade-off. Open questions have no set answers, providing detail that requires coding. Closed questions offer fixed options; while ranked or agreement scales produce ordinal data, non-ranked choices like yes/no or categories yield nominal data.

This is the prediction a study is set up to test, and it must be operationalised, saying exactly what will be measured and how it will be scored, so a marker can see what would count as support. Pair it with the null, which claims that any difference or relationship found is down to chance. Choose the one tailed version, naming the direction, only when earlier research points one way, and the two tailed version when findings conflict or the area is new, because a direction predicted on no evidence cannot be defended. The wording follows the design, a difference between conditions for an experiment, a relationship between two measured variables for a correlation, an association for a test of categorical frequencies. The decision comes at the end, when the calculated value is set against the critical value at the five per cent level.

A sample is the group actually studied, drawn from the target population the researcher wants to say something about, and the reason for choosing it carefully is generalisation. Examiners expect the target population named, the list the researcher drew from named, and the bias each method leaves behind explained. Milgram recruited forty American men through a newspaper advertisement, so his obedience rate describes self selecting volunteers rather than people in general, and Asch tested male college students who happened to be available to him. Sherif and colleagues went the other way, screening and matching twenty two boys so that the two groups began alike. Judge any sample by who is missing from it, and connect that to androcentrism, ethnocentrism and the limits of a nomothetic claim.

Random selection gives every member of the list an equal chance, usually by numbering that list and drawing with a random number generator, so bias is removed in principle although a complete list of a population is rarely available. Stratified selection identifies subgroups such as age bands or sex, works out the proportion of the population each forms, then samples randomly inside each stratum, which reproduces the population structure but needs that structure known in advance. Volunteer or self selecting samples come from advertisements, as in Milgram and later in Burger, and attract people who are compliant, curious and unusually cooperative. Opportunity sampling takes whoever is available, which is quick and cheap but hands the choice to the researcher and over represents students.

Numerical data, such as the percentage of participants who continued to the highest switch, allow statistical testing, replication and clear graphical summary, at the cost of stripping out the reasons behind the number. Data recorded as words or images, such as interview transcripts, diary entries and observational field notes, preserve meaning and context but resist summary and depend on whoever interprets them. Milgram reported both, giving obedience rates alongside descriptions of participants sweating, laughing nervously and pleading to stop, and the pair says more than either alone. The form of the response decides the type, not the method, since an interview or a questionnaire can yield either. Tie the distinction to reductionism against holism and to nomothetic laws against idiographic study of the case.

Find the mean by totalling scores and dividing by the count, the median by ordering scores and taking the middle value, and the mode by finding the most frequent score. The mean uses every score but is skewed by extremes; the median better summarises skewed sets, while the mode suits categorical data. Frequency tables organise raw data by listing each score alongside how often it occurs, aiding the calculation of central tendency and chart construction. Spread matters equally: the range is the highest minus the lowest score, resting on two values, while standard deviation measures the average distance of all scores from the mean. Bar charts display categorical data with gaps between bars.

This method turns transcripts, diaries or open questionnaire answers into patterns of meaning. Braun and Clarke set out six phases examiners expect in order: becoming familiar with the material by repeated reading, generating initial codes for anything of interest, collating codes into candidate themes, reviewing those themes against the whole data set, defining and naming them, and writing up with verbatim quotations as evidence. Coding may be inductive, letting themes rise out of the data, or deductive, testing themes taken from existing theory, and interpretation may stay semantic at the surface meaning or go latent to the assumptions beneath it. Credibility comes from a second coder checking agreement, from an audit trail of decisions, and from taking themes back to participants.

These are enforceable professional expectations rather than laws, and the examinable skill is applying them to a study described in the question. Informed consent means agreeing on the basis of a true account of the procedure, which deception removes, so presumptive consent, prior general consent and a thorough debriefing are the usual repairs. The right to withdraw must be genuine and must extend to withdrawing data afterwards, protection from harm means risk no greater than ordinary life, and confidentiality means data nobody can trace back to a person. Milgram's participants were deceived about the shocks and showed real distress, Sherif's boys never knew a study was happening, and Burger's later work shows how a design can be rebuilt to modern standards.

The code is built on four principles: respect for the dignity and autonomy of the person, competence to attempt the work in hand, responsibility for participants and for the standing of the discipline, and integrity in reporting honestly. The version published in that year replaced a list of rules with principles to be reasoned about, and later revisions have kept the structure. Managing risk asks whether taking part carries harm greater than that met in ordinary life, and calls for a written assessment, extra safeguards for children, vulnerable adults and sensitive topics, monitoring for distress with a rule for stopping, and attention to the safety of the researcher in fieldwork. Approval by an ethics committee and a named person to receive concerns are part of the same system.

This is the only method that supports a claim about cause, because the researcher changes one thing and holds everything else as still as the setting allows. Teach the decisions in the order they are made: operationalise the variable that is manipulated and the one that is measured, choose the setting, decide whether each person does one condition or all of them, then control the rest with standardised instructions, random allocation and counterbalanced order. Every choice trades control against realism. Loftus and Palmer (1974) altered a single verb under laboratory conditions and could be confident the verb caused the change in speed estimates, while Hofling and colleagues (1966) telephoned nurses on real wards and learned about obedience in a genuine hierarchy at the cost of knowing what else was happening.

The difference lies in where the study happens, not in how carefully it is done. Under laboratory conditions the researcher controls the setting as well as the manipulation, so the procedure can be standardised and repeated exactly. In a natural setting the manipulation is made where participants already were, so behaviour is more typical but the noise cannot be removed. Running either one means writing an operationalised hypothesis, piloting it, sampling and allocating participants, scripting instructions that every participant hears in the same words, briefing and debriefing, and deciding in advance how the measure will be scored. Piliavin and colleagues (1969) staged collapses on a moving underground train, capturing helping behaviour nobody would report honestly on a questionnaire, yet they could not gain consent or control who was in the carriage.

Everything in a results section rests on getting these two right. One is the thing the researcher changes and predicts will have an effect; the other is the outcome that is measured and expected to change with it. Anything else that could move the outcome is extraneous, and if it changes systematically alongside the conditions it becomes a confounding variable that makes the finding uninterpretable. Operationalising is where the marks are won and lost, because aggression cannot be measured but the number of imitated acts in ten minutes can, and memory cannot be measured but the number of words recalled in the correct order can. Baddeley varied the type of word list and counted words recalled in order; Loftus and Palmer varied a single verb and recorded estimated speed in miles per hour.

A hypothesis is a testable prediction written before any data are collected, and the marks go to precision rather than to ideas. A prediction of a difference or a relationship must name both variables in measurable terms and, when it is directional, say which way the effect will go; it should only be directional where earlier findings or theory justify the direction, and otherwise the non-directional form is safer. The rival statement says any difference observed is due to chance alone, and it is the one actually tested, because a statistical test gives the probability of the obtained result if that statement were true. At the usual five per cent level a sufficiently unlikely result allows it to be rejected and the prediction retained.

A hypothesis is a testable prediction written before any data are collected, and the form it takes commits the researcher to a particular reading of the statistics. Predicting that participants who see a violent film will name fewer items than controls fixes the direction; predicting only that the two groups will differ leaves the prediction open at both ends. That choice changes the numbers, because with the same sample size and the same level of significance the whole rejection region sits in one tail of the distribution rather than being split across two, so the critical value is easier to reach. Loftus and Palmer had grounds for a directional prediction, since leading questions were already known to bias recall, whereas a genuinely new manipulation usually justifies leaving the prediction open. Every prediction must also be operationalised and paired with a null hypothesis.

The design a study uses decides which threat to validity it must defend against. Testing the same people in every condition removes individual differences and needs fewer participants, but it invites practice and fatigue effects and makes the aim easier to guess. Using different people in each condition removes those order effects but lets individual differences load onto one condition, which is why random allocation matters. Matching pairs on a variable known to affect the outcome, as Bandura, Ross and Ross did with nursery children matched on aggression ratings before allocation to model conditions, buys part of the control of a repeated design without the order effects, at the cost of a pre-test and lost data when one member of a pair drops out. Loftus and Palmer used separate groups for their five verb conditions, because nobody sees the same crash film for the first time twice.

To operationalise is to write a variable as the exact procedure that will produce the numbers, so aggression becomes a count of imitative mallet strikes in a fixed observation period, as in Bandura's work, and memory for an accident becomes a speed estimate in miles per hour, as in Loftus and Palmer. The gain is replicability and agreement between observers; the loss is that a narrow measure may not capture the whole construct, which is the reductionism criticism examiners reward. Anything besides the independent variable that could nudge the outcome is a nuisance, and most of it is random noise spread evenly across conditions. It becomes a genuine rival explanation only when it changes systematically with the conditions, for instance when the quiet group is tested in the morning and the noisy group after lunch, because the results then have two explanations that cannot be separated.

Testing the same person twice buys control over individual differences and then hands back a new problem, because performance in the second condition is never clean: practice can lift it and fatigue or boredom can depress it. Counterbalancing does not remove that contamination, it spreads it evenly, either by splitting the sample so half meet the first condition first and half meet the second condition first, or by running each participant through a sequence in which each condition sits equally early and late. Randomisation attacks the same family of problems by letting chance decide the order of trials, the order of items in a word list or which stimulus a participant meets, which stops a researcher unconsciously putting the easy trials first. Peterson and Peterson varied their retention intervals for this reason. None of this is the same thing as random sampling or random allocation.

One family of nuisance variables belongs to the setting and the procedure: the time of day, the noise level, the clarity of the instructions, the manner of the researcher. These are handled by standardising everything except the manipulation. The other family belongs to the people: age, prior knowledge, mood, motivation, memory span. These are handled by allocating participants at random, by matching them on the characteristic that matters or by testing the same people throughout. Milgram's variations show why the distinction reaches beyond design: moving the study from Yale to a run-down office and bringing the learner into the same room both cut obedience sharply, which supports a situational rather than a dispositional account of obeying. In any study using separate groups, uncontrolled differences between people are the standard rival explanation for a difference in scores.

These words are not interchangeable, and a measure can be perfectly consistent and consistently wrong. Freedom from the researcher's interpretation comes from standardised measures, clear categories and blind procedures. Consistency is checked by correlating two observers' records of the same behaviour, by retesting people after an interval, or by splitting a questionnaire in half, and is improved by training observers and tightening the categories. Then come the truth questions: whether the manipulation really produced the change rather than a confounding variable, whether a score forecasts a later outcome as an aptitude test should forecast job performance, and whether the finding survives outside the setting. That last is the charge against Loftus and Palmer, since Yuille and Cutshall found witnesses of a real shooting stayed accurate months later and resisted leading questions.

Participants are not passive. They build their own theory of what a study is about and adjust to it, either helping the researcher, deliberately spoiling the data or managing how they appear, which is social desirability. Orne showed how strong the participant role is when volunteers went on adding columns of figures and tearing up their answers for hours, because the situation told them that was what taking part required. The bias can also run the other way, from the researcher. Rosenthal and Fode found that students told their rats had been bred to be maze-bright recorded better performance than students given the same animals described as maze-dull, and the classroom study with Jacobson found children labelled as likely to bloom gained more on retesting. The remedies are blind procedures, standardised instructions and cover stories, the last of which buys control at an ethical price.

Choosing a summary depends on the level of measurement. The mean uses every score but is dragged by extreme values, so a skewed set is better described by the median. The mode is the only option for categorical data. Frequency tables organise raw data by tallying how often each value occurs. Spread completes the picture: the range uses only the highest and lowest scores. Standard deviation reflects the whole set; to calculate it, subtract the mean from each score, square these differences, sum them, divide by the number of scores minus one, and take the square root. Percentages allow fair comparison of unequal groups. Graphs follow data type: bar graphs for discrete categories and histograms with touching bars for continuous data.

The test is chosen by the design and the data. Two separate groups of people compared on ordinal or better data call for Mann-Whitney; the same people or matched pairs measured twice call for Wilcoxon signed ranks. Both are distribution free, so they survive a skewed set of scores, where a long tail drags the mean away from the middle value. Both are also read against the table in the way that surprises students, because the calculated value has to be equal to or smaller than the tabled value before the null hypothesis can be rejected. Finding the right tabled value needs the sample sizes, whether the prediction was directional, and the level of significance, conventionally five in one hundred. Rejecting a true null hypothesis is a Type I error, retaining a false one is a Type II error, and lowering the risk of either raises the risk of the other.

A single patient, or a very small group, studied in depth over years with interviews, observation, psychometric testing and later brain imaging, is how cognitive psychology reached conclusions no experiment could ethically create. Henry Molaison had medial temporal tissue including most of the hippocampus removed to control intractable epilepsy, and Scoville and Milner (1957) found he could hold a normal digit span and score normally on intelligence tests yet form almost no new lasting memories, while Milner recorded his mirror drawing improving day after day although he denied having done the task before. Corkin and colleagues later used imaging to fix what had actually been removed. Marks come from weighing that evidence, not from retelling his life.

Nothing is manipulated here: two variables are measured as they already stand in the same people and the analysis asks whether they vary together, so the pair are called co-variables and there is no independent or dependent variable. A positive relationship means high scores on one go with high scores on the other, a negative one means high scores go with low, and a near zero result means the pair move independently. The scatter diagram plots one co-variable on each axis with one point per participant, and the tighter the points lie to a line of best fit the stronger the relationship. Because neither variable was controlled, a relationship cannot show direction of cause, and an untested third variable may drive both, which is why the link between testosterone and violent offending that Dabbs and colleagues reported is open to the reading that dominance raises the hormone.

Correlational data analysis involves interpreting scatter diagrams and using Spearman’s rho to test statistical significance. Rho is used for relationships between co-variables with related pairs of at least ordinal data. To calculate, rank each co-variable separately, find the difference between paired ranks, square and sum these differences, and apply the formula. The observed value is compared to a critical value based on N, significance level, and hypothesis direction. For rho, the absolute observed value (ignoring any minus sign) must be equal to or greater than the critical value to reject the null hypothesis. Correlations use alternate and null hypotheses. Experiments use experimental and null hypotheses, manipulating an IV to measure a DV, utilising control groups, random allocation, and specific sampling methods.

Scanning splits three ways and the examiner wants the split. Computerised axial tomography stacks X-ray slices into a still picture of structure, so it shows a tumour or a lesion but never activity. Positron emission tomography injects a radioactive glucose tracer and maps where the brain is burning fuel, which is functional but slow and delivers a radiation dose. Functional magnetic resonance imaging follows the blood oxygen level dependent signal, giving detail close to a millimetre with no radiation, though the picture lags the neural event by a few seconds. Heritability is asked separately: twin designs compare identical with non-identical concordance, adoption designs pull biological parentage away from rearing, and neither on its own fully separates gene from environment.

Observational research gathers qualitative data, such as descriptive notes of behaviour, and quantitative data, such as frequencies through tallying. Event sampling records every occurrence of a target behaviour, while time sampling records behaviour at fixed intervals. Observations vary by type: participant (observer joins the group), non-participant (observer remains separate), structured (controlled setting or tasks), naturalistic (unaltered everyday setting), overt (participants know they are observed), and covert (participants are unaware). Content analysis applies a coding system to existing media, converting qualitative material into quantitative data by counting category instances.

Animals are used where the control required would be impossible or unacceptable with people: Pavlov timed salivation to a signal, Skinner's chamber isolated one schedule at a time, and Seligman and Maier showed that dogs given uncontrollable shock later failed to escape when escape was available. The legal framework in Britain runs through the Animals (Scientific Procedures) Act of 1986, which requires three Home Office licences at once, covering the project, the individual researcher and the premises, and which grants them only after a cost-benefit judgement that the likely gain outweighs the suffering. Inspectors enforce it, and the three Rs of replacement, reduction and refinement shape what a licensed procedure may actually do.

Choosing the test is a three-part decision: the data are frequency counts in categories, so nominal; the design is independent groups with each participant appearing in one cell only; and the hypothesis predicts a difference or an association. That combination points to chi-squared and to nothing else. An expected frequency comes from the row total multiplied by the column total and divided by the grand total, each cell then contributes the squared difference between observed and expected divided by expected, and degrees of freedom come from rows minus one multiplied by columns minus one. Thematic analysis is the qualitative counterpart, coding transcripts, grouping codes into themes, checking those themes back against the data and reporting them with verbatim quotation.

Science is treated here as a set of practices rather than as a subject matter, and the case for psychology rests on its use of them: standardised procedures that allow a study to be repeated, measures checked for consistency, controls and control groups that protect the claim the manipulation caused the change, testable hypotheses drawn from theory, and conclusions taken only from observable data. Popper's demand that a theory forbid something is the sharpest evaluative tool, because it separates behaviourist predictions from psychodynamic claims that absorb any outcome. Against the case stand Kuhn's argument that the discipline has no single shared paradigm, the reactivity of participants who work out the aim, and a large replication project that reproduced only around a third of published findings.

Practitioner psychologist is a protected title in the United Kingdom, so a clinical psychologist must appear on the register and must meet the published standards of conduct, performance and ethics. Those standards require a practitioner to promote and protect the interests of service users and carers, to communicate appropriately and effectively, to work only within the limits of their knowledge and skills and refer on where necessary, to keep information confidential and share it only with consent or lawful reason, to manage risk, to report concerns about safety, to be open and honest when something goes wrong, and to keep full and accurate records. Breach leads to fitness to practise proceedings, and the ultimate sanction is removal from the register and loss of the right to practise.

Each design answers a different question. Following one sample across years shows the course of a disorder and removes participant differences between time points, but loses people through attrition and outruns most funding. Comparing age groups at a single moment is quick and cheap yet confounds age with cohort, because people born in different decades met different diagnostic fashions. Comparing societies tests whether a category is universal or an imposed etic, which is what the World Health Organization work on outcome in several countries put in question when recovery looked better outside industrialised nations. Pooling many studies statistically raises power and settles disagreements, yet inherits publication bias and mixes trials that measured different things. Primary data fits the question exactly; secondary data is cheap and already sitting in records.

A case study takes one person or one small group and gathers many kinds of data about them over time, typically interviews, observation, clinical records and sometimes test scores, and it is idiographic because it aims at understanding this instance rather than estimating a population value. Lavarenne and colleagues reported on a therapy group for people with psychosis whose sense of a boundary between self and others was fragile, and described how the group itself came to act as a containing shell, the shared structure of the sessions holding anxiety that individual members could not hold alone. The approach suits rare, sensitive or unrepeatable situations and generates hypotheses, but the findings rest on the researcher interpretation, cannot be replicated and generalise no further than the group studied.

Interviews run from a fixed schedule read in the same order to every participant, through a semi-structured form carrying set topics and the freedom to follow an answer, to an unstructured conversation guided only by an aim. Vallentine and colleagues ran psycho-educational sessions for male patients detained in a high-security hospital and used questionnaires and interviews to find out what the men took from them; participants reported a better grasp of their illness and their medication and valued discovering that others shared the experience, while measured gains in factual knowledge were more modest. Interviews give depth and let a person explain in their own words, but social desirability, the power gap between a detained patient and a clinician, and the interpretation of transcripts all threaten validity.

Choose a test from three questions: difference or relationship, what design, and what level of measurement. Frequencies in independent categories take chi-squared, ordinal scores from two separate groups take Mann-Whitney, ordinal scores from the same or matched people take Wilcoxon, and paired scores tested for association take Spearman. The direction of the critical value comparison catches people out, because chi-squared and Spearman must equal or exceed the critical value while Mann-Whitney and Wilcoxon must be equal to or smaller than it. Qualitative data are coded rather than counted: thematic analysis works through transcripts to recurring themes, while grounded theory builds categories by constant comparison until saturation and lets the theory emerge from the data rather than testing one set in advance.

Method choice is the examinable decision. A laboratory study gives control: Loftus and Palmer (1974) manipulated a single word, held the film, the room and every other question constant, and could therefore claim the verb caused the change in estimates, at the price of a filmed collision nobody cared about. Moving the manipulation outdoors keeps the independent variable but loses that control, as when Maass and Kohnken approached students holding either a syringe or a pen and found recall of the person's face was worse in the threatening condition, which buys realism and gives up command of who walks past and in what mood. A case study takes one witness, one offender or one crime and goes deep, as Yuille and Cutshall did with the witnesses to a real shooting.

Sampling is examined as a procedure you could carry out, not as four definitions. Random selection needs a complete list of the target population and a device that gives every name an equal chance, such as a random number generator applied to a numbered roll. Stratified sampling first identifies the subgroups that matter, offence type or age band, works out what proportion of the population each holds, and then samples randomly within each subgroup in that proportion, so a wing where three quarters are sentenced for violence yields a sample that is three quarters violent. Volunteer samples come from an advertisement and gather the willing and the curious. Opportunity samples take whoever is present and agrees, which is fast, and is why so much offender research rests on one institution.

These five words are marked as five separate judgements. Reliability is consistency, checked by test retest correlation on the same measure, by split half on a personality questionnaire, and by inter rater agreement when two observers code aggression on a prison wing. Validity asks whether the measure captures what it claims, internally when confounds are ruled out and externally when the finding travels to other people, settings and times. Objectivity is the removal of the researcher's expectations through standardised procedure, blind scoring and placebo control. Credibility is whether an account is believable, raised by checking interview data against another source, as the Vancouver witness study did with police records. Ethics runs through all of it, because offenders, jurors and distressed witnesses are among the least free participants psychology studies.

Choosing an inferential test requires checking if the hypothesis predicts a difference or correlation, the experimental design, and the level of measurement. Nominal data from an unrelated design point to chi-squared; independent groups with ordinal data point to Mann-Whitney; repeated measures with ordinal data point to Wilcoxon; a correlation between ordinal variables points to Spearman. Significance involves comparing observed and critical values: chi-squared and Spearman must reach or exceed the critical value, while Mann-Whitney and Wilcoxon must be equal to or lower. Quantitative analysis also involves frequency tables to tally scores, measures of central tendency, dispersion (range, standard deviation), and meta-analysis. Qualitative data uses thematic analysis and grounded theory.

Two different documents doing two different jobs, and the examiner rewards the distinction. The professional body sets out four principles, respect, competence, responsibility and integrity, which between them cover informed consent, confidentiality, the right to withdraw, working within the limits of your training, protection from harm, debriefing and honesty about findings. Risk management is the prospective half of that work: before any data are collected the researcher assesses what could go wrong for participants and for the researcher, records how each risk is reduced, and plans screening, distress protocols and referral. The statutory regulator instead governs practitioners, who must be registered to use the protected title, keep records, work inside their scope of practice and base formulation and intervention on evidence, and who can be removed from the register altogether.

Watching records what children do rather than what they say about it, which matters when the participants are too young to report on themselves. The design choices carry the marks: behavioural categories operationalised in advance, then event sampling or time sampling, with a tally of each occurrence producing frequencies, alongside narrative field notes that give the qualitative detail a tally cannot. Ainsworth and Bell (1970) is the model, with trained coders behind a one-way screen scoring proximity seeking, contact maintaining, avoidance and resistance in fifteen second intervals. Bandura's Bobo doll work used the same machinery and reported close agreement between independent coders, which is how reliability is shown.

Self-report gathers what watching cannot see, which in child psychology usually means parents answering for young children and older children answering for themselves. A fixed schedule of closed items gives comparable, easily analysed numbers but forces replies into the researcher's categories; a conversation around open questions gives rich accounts at the cost of consistency and of analysis time. The middle option is the Adult Attachment Interview, where set questions about childhood are followed up flexibly and the classification comes from how coherently the adult tells the story. Sampling shapes the result too, since Hazan and Shaver's love quiz recruited whoever chose to reply to a newspaper, which is cheap and self-selecting.

Comparing societies asks whether a finding travels, and attachment is the standard test case because one procedure has been run in dozens of countries. van IJzendoorn and Kroonenberg (1988) pooled thirty-two such samples from eight countries and found secure the majority classification everywhere, avoidant most frequent in West German samples, resistant most frequent in Japanese and Israeli ones, and variation between samples inside a country about one and a half times the variation between countries. The design question follows: comparing different age groups at one time is quick but confounds age with cohort, while following the same children for years measures development itself at the cost of time and of participants dropping out.

A child cannot give legally valid consent, so research runs on a guardian's permission plus the child's own assent, sought in words the child understands and treated as withdrawable through distress or reluctance rather than only through speech. The United Nations Convention adds duties a psychologist has to honour, above all that the child's best interests come first and that children may express views on matters affecting them. That is where the tension the specification cares about comes from, because keeping children out of studies protects them from harm while also silencing them in decisions about day care, schooling and treatment that are made on adult testimony alone.

Choosing a test means answering three questions in order: difference or relationship, related or unrelated participants, and level of measurement. A relationship with ordinal data gives Spearman's rank correlation; a difference between unrelated groups with ordinal data gives Mann-Whitney; a difference with the same or matched participants gives Wilcoxon signed ranks; and counts of people in nominal categories with independent participants give chi-squared. The comparison with the critical value then splits into two families, and the sign is what students reverse: for Spearman and chi-squared the observed value must be equal to or larger than the critical value, while for Mann-Whitney and Wilcoxon it must be equal to or smaller.

Animal work buys control that no human study can offer. The strain, the diet, the exact dose, the entire drug history and invasive surgery are all in the experimenter's hands, a generation passes in months, and procedures such as implanting electrodes or lesioning a pathway could never be done to a person. The standard paradigms are intravenous self-administration, in which an animal presses a lever for an infusion, conditioned place preference, in which the animal is later tested for the chamber it associates with the drug, and intracranial self-stimulation of the reward pathway. The cost falls on an animal that cannot consent and may be made dependent, so British law and the recognised ethical frameworks work by cost and benefit rather than by prohibition.

Two methods, and the marks are in the trade-off between them rather than in two separate descriptions. A controlled experiment, usually double blind and placebo controlled with random allocation, separates the pharmacology from expectation and permits a causal claim, at the price of brief exposures, modest doses and volunteers who are rarely much like a long term dependent user. Interviews and longitudinal cohort studies follow real users over years and capture how a drug career begins, is funded and ends, at the price of attrition, retrospective recall and social desirability. Ethics sits on top of both, because participants may be dependent, criminalised or hoping that the study is treatment, which makes freely given consent the hardest principle of all to satisfy.

Comparing substance use between cultures is the nearest psychology gets to a natural experiment on inheritance and environment, because gene frequencies and drinking customs move at very different speeds. Populations in which a variant of the aldehyde dehydrogenase enzyme is common show flushing, nausea and headache after alcohol and lower rates of dependence, which looks biological, and yet consumption in those same populations has risen and fallen sharply across decades in which the gene frequencies did not change at all, which is the environment talking. Law, price, availability, religion and what counts locally as a normal drink all differ, and so does the way the question is asked, so an instrument written in one culture and exported to another can manufacture the difference it then reports.

Choosing an inferential test is a three-part decision: is the hypothesis about a difference or relationship, are scores from separate or the same people, and what is the level of measurement. Nominal counts point to chi squared, ordinal scores from independent groups to Mann-Whitney, matched participants to Wilcoxon, and ranked variables to Spearman. Chi squared and Spearman are significant when the calculated value (using the absolute value for negative Spearman correlations) is equal to or larger than the critical value; Mann-Whitney and Wilcoxon when equal to or smaller. Quantitative data is also analysed using frequency tables, central tendency and dispersion. Qualitative data is coded into themes or built upwards into a theory.

Quantitative work counts something and gives numbers you can average and test; qualitative work records what people say or do in their own words. Evidence is primary when the researcher gathered it first hand for the current aim, and secondary when it already existed for some other purpose, such as hospital records, crime figures or another team's published results. Milgram measured obedience as the maximum voltage delivered, which is a number, but also logged sweating, stuttering and nervous laughter, which is the descriptive half of the same study. Van IJzendoorn and Kroonenberg pooled other researchers' Strange Situation results, so every figure they analysed was second hand.

A sample stands in for a target population, and each method trades representativeness against practicality. In a random method every member has a calculable and equal chance, usually by numbering a full list and drawing with a random number generator, so selection bias is removed but the list rarely exists. A stratified approach identifies subgroups such as age or sex, works out the proportion each forms of the population and then draws randomly inside each layer, which is the most representative and the most laborious. Self selected participants answer an advertisement, as Milgram's did through a newspaper notice, and tend to be unusually obliging. Asch simply used the students available to him, which is why his conformity work rests on young American men.

Design is the question of how participants are allocated to conditions, and the answer fixes both the threats to validity and the test used later. When separate groups each take one condition there are no order effects, but individual differences can swamp the manipulation unless allocation to conditions is random. When the same people take every condition those differences disappear and fewer participants are needed, yet practice, fatigue and guessing the aim now bite, so the order is counterbalanced. Pairing people on a relevant variable and splitting each pair across conditions keeps the groups separate while controlling the variable that matters, which is what Bandura, Ross and Ross did by rating children for aggression before allocating them, though matching is slow and never exact.

A hypothesis is a testable, fully operationalised prediction, and the analysis depends on how it is worded. The one that is actually tested states that there is no difference or no relationship and that anything observed is down to chance; it is retained or rejected against a critical value. The prediction of a difference or a relationship is called experimental only where an independent variable has been manipulated, and alternate elsewhere, for instance in correlational or questionnaire work. Predicting which way the result will fall is justified only when earlier findings point one way, and the neutral wording is chosen when the evidence is mixed or contradictory. That choice then decides whether a one-tailed or a two-tailed critical value is read from the table.

Self report asks people to tell you about themselves, and the format decides what comes back. Fixed response items, including ranked or Likert style scales, produce numbers that are quick to compare but force respondents into boxes and yield ordinal rather than interval scores. Items that invite people to write freely produce detail that must be coded before it can be summarised. A fixed script replicates well and lets different interviewers be compared; following the respondent digs deeper but is hard to repeat; a schedule with probes sits between the two. Adorno and colleagues measured the authoritarian personality with a fixed response scale, and the criticism that stuck was acquiescence, because every item was worded so that agreeing pushed the score up.

In an experiment the researcher manipulates one variable and measures its effect on another while holding the rest steady, which is what licenses a causal claim. Doing this in a controlled setting with standardised procedures gives precise measurement and easy replication, but the task can be artificial and participants know they are being studied. Doing it in a real setting gives more natural behaviour and weaker demand characteristics, at the cost of looser control and, usually, of informed consent. Loftus and Palmer changed a single verb in a question and moved mean speed estimates from around forty miles per hour for smashed down to around thirty for contacted, while Piliavin and colleagues staged collapses on a moving subway train and timed how quickly passengers helped.

Watching behaviour only becomes evidence once there is a system for recording it. Categories must be observable, exhaustive and mutually exclusive so that two people counting the same episode agree, and agreement is checked by correlating their totals, with a coefficient of about 0.8 treated as acceptable. Counting every instance of a category suits behaviour that is rare, whereas recording what is happening at fixed intervals suits behaviour that is more or less continuous but misses whatever falls between the checks. A coding scheme in a controlled setting, as in Ainsworth's Strange Situation, buys reliability, while watching secretly in a natural setting buys realism at the cost of consent and privacy. Bandura's team scored imitative aggression in short fixed intervals from behind a one-way mirror.

Each technique is examined for what it can and cannot establish. Comparing identical with non-identical pairs holds the family environment constant while varying genetic similarity. Gottesman's schizophrenia twin data put identical concordance near half against a sixth for non-identical pairs, supporting genetic influence but ruling out inevitability. Adoption studies separate biological from rearing parents. Animal experiments permit control that would be unethical with people. Case studies give depth on rare patients at the cost of generalisability. Scanning methods like fMRI observe brain activity. Cross-cultural research compares societies to identify universal traits, and meta-analysis pools data across studies to find overall effects.

Control means making sure the manipulated variable is the only thing that changed. Any nuisance variable that might affect the measure is a threat; it becomes genuinely damaging once it varies systematically with the conditions, because the result can then no longer be attributed to the manipulation. Differences between people, such as memory span or age, and features of the setting, such as noise or time of day, are the two usual sources. Testing the same people twice adds practice and fatigue, which is spread evenly by having half take the conditions in reverse order. Orne showed that participants read cues and act on what they think is wanted, and Rosenthal showed that a researcher's expectations leak into the data, which is why standardised instructions and double blind procedures exist.

Central tendency and dispersion are chosen to fit the data, not by preference: the mean uses every score and suits interval data, the median resists extreme scores and suits ordinal or lopsided sets, and the mode is the only option for named categories. A bar chart leaves gaps because its categories are discrete, a histogram has touching bars because the variable is continuous, and a scatter diagram plots two co-variables to show a relationship. Standard deviation reports the average distance of scores from the mean, so a small value means a tightly clustered group and a mean that represents it well. Loftus and Palmer reported mean speed estimates of 40.8 mph after the verb smashed and 31.8 mph after contacted, which shows how a comparison of two means carries both direction and size.

Test choice follows three questions asked together: is the study looking for a difference or a relationship, is the design independent or repeated, and what is the level of measurement. Data in named categories point to chi squared, ranked data from independent groups to Mann-Whitney, ranked data from repeated measures or matched pairs to Wilcoxon, and a relationship between two sets of ranks to Spearman. The decision rule then splits, because the observed value must be equal to or smaller than the critical value for Mann-Whitney and Wilcoxon, and must reach or exceed it for Spearman and for chi squared. The critical value itself is read against the number of participants or the degrees of freedom, the number of tails and the chosen level of significance, and a result that passes the comparison allows the null hypothesis to be rejected.

These are the quality checks a study is judged by. Internal validity asks whether the independent variable, rather than a confounding variable or a demand characteristic, produced the change measured. Predictive validity asks whether a measure forecasts later behaviour, as an aptitude test should forecast performance in the job. Ecological validity asks whether the finding survives outside the setting and the task used. Reliability is consistency, shown by test retest, split half or agreement between raters using the same behavioural categories. Generalisability rests on the sample, which is why Asch tested only male American students and Milgram only men. Objectivity means a conclusion the data force on any observer, researcher bias is what the researcher adds to it, and credibility is the qualitative counterpart, built through triangulation and checking with participants.

Both methods turn words into findings, but they start from opposite ends. Thematic analysis, set out by Braun and Clarke, moves through six phases: familiarisation with the transcripts, generating initial codes, searching for themes, reviewing them against the whole data set, defining and naming them, and writing them up with quotations as evidence. Themes may be inductive, driven by the data, or deductive, brought to it. Grounded theory, developed by Glaser and Strauss, builds a theory from data rather than testing one, using open, axial and selective coding, constant comparison of each new case with earlier ones, and theoretical sampling that continues until saturation, the point where fresh interviews add no new categories. Both give rich detail about meaning, and both are attacked for the interpretation the researcher brings to it.

A journal article follows a fixed order so that a reader can find anything quickly and a rival can repeat the work. The abstract summarises aim, method, results and conclusion in a paragraph. The introduction funnels from previous research to a gap, then to the aim and to hypotheses that are operationalised and stated as directional or not. The method gives design, participants and sampling, materials, procedure and ethics in enough detail to replicate. The results report descriptive statistics first and then the inferential test with its observed value, critical value and significance. The discussion interprets, relates the findings back to earlier work, admits limitations and suggests what to do next. Before any of this is printed, other experts in the field read the manuscript anonymously and recommend acceptance, revision or rejection.

The British Psychological Society organises its code around four principles, respect, competence, responsibility and integrity, and the rules follow from them: informed consent, no deception where an alternative exists, the right to withdraw at any point, confidentiality, protection from harm and a full debrief. Milgram obtained agreement to a study of memory and punishment, not to being ordered to give what participants believed were four hundred and fifty volt shocks, and his observations record trembling and sweating, so his defence rested on the debrief and on the follow up in which most said they were glad to have taken part. Researchers meet these problems with presumptive or retrospective consent, with ethics committees weighing costs against benefits, and with a risk assessment covering hazards to participants and to the researcher before data collection starts.

Animal research in Britain runs on licences, not on a professional code. Protection covers vertebrates and cephalopods, and any regulated procedure needs three licences at once: a personal licence held by the individual who carries out the work, a project licence for the programme of research, and a licence for the place where it happens. Home Office inspectors enforce them, and approval rests on a judgement that the likely benefit outweighs the suffering, with extra justification for cats, dogs, horses and primates. Researchers are expected to work to replacement, reduction and refinement, using a non animal method where one exists, the smallest adequate number, and procedures that limit pain, with housing, anaesthesia and humane killing specified. Harlow reared rhesus monkeys with wire and cloth surrogates, which shows both what such work reveals and why the rules exist.