Research Methods in Cognitive Psychology and Cognitive Science

Last reviewed on October 2, 2026.

How do you study something you cannot see? Cognitive scientists cannot observe a memory or a decision directly, so they infer mental processes from what people do, from what their brains do, and from how well computational models reproduce both. Because the field is interdisciplinary, its toolkit borrows from experimental psychology, neuroscience, computer science and linguistics. This guide explains each of the main research methods used in cognitive psychology and cognitive science, with classic experiments and the strengths and limits of each.

Research Methods at a Glance

Behavioural Experiments: The Core of the Cognitive Approach

Behavioural experiments are the oldest and still the most widely used method in cognitive psychology. The researcher manipulates an independent variable (for example, the delay before recall) and measures a dependent variable (how many items are remembered), holding everything else constant. If performance changes reliably with the manipulation, the change tells us something about the hidden process.

Classic experiments for the cognitive approach

• The Stroop effect (Stroop, 1935). People name the ink colour of colour words. When the word conflicts with the ink (“RED” printed in blue), naming is slower and more error-prone. The experiment shows that reading is so automatic that it cannot easily be switched off, and it remains a standard measure of interference control.
• Iconic memory (Sperling, 1960). Participants saw a grid of 12 letters for 50 milliseconds. Asked to report all of them, they managed about four. But when a tone straight after the display told them which single row to report, they could usually report most of that row – implying that most of the grid was briefly available in a fleeting visual store that fades within about a second.
• Short-term forgetting (Peterson & Peterson, 1959). Participants heard a three-consonant trigram, then counted backwards in threes to prevent rehearsal. Recall fell steeply, to around 10% after 18 seconds, showing how quickly unrehearsed information is lost from short-term memory.
• Mental rotation (Shepard & Metzler, 1971). Participants decided whether two drawings of 3D block shapes were the same object rotated or mirror images. Response time increased in a straight line with the angle between them, as if people were rotating a mental image at a constant speed – strong evidence for analogue mental imagery.
• Leading questions and eyewitness memory (Loftus & Palmer, 1974). After watching films of car crashes, participants asked how fast the cars were going when they “smashed into” each other gave higher speed estimates (roughly 41 mph) than those asked about cars that “contacted” each other (roughly 32 mph). In a follow-up, the “smashed” group was more likely to falsely remember broken glass a week later. Memory is reconstructive.
• Working memory (Baddeley & Hitch, 1974). Using dual-task experiments, participants held strings of digits in mind while reasoning or comprehending sentences. Performance slowed only modestly, which is hard to explain if short-term memory were a single store – leading to the multi-component model of working memory.
• Attentional cueing (Posner, 1980). A cue indicates where a target will probably appear. People respond faster when the cue is valid and slower when it is invalid, even without moving their eyes, showing that attention can be shifted covertly like a spotlight.

Many of these studies are discussed in more detail in our articles on memory and learning and attention and perception.

Reaction time and accuracy

The two workhorse measures are reaction time (how long a response takes, usually in milliseconds) and accuracy (the proportion of correct responses). Reaction time goes back to Franciscus Donders, who in 1868 subtracted the time for a simple response from the time for a choice response to estimate how long a decision takes. Today researchers also look at the whole distribution of response times and at the trade-off between speed and accuracy, since people can go faster by accepting more errors. Formal models such as the drift-diffusion model (Ratcliff, 1978) separate how quickly evidence accumulates from how cautious a person is.

Within-subjects and between-subjects designs

In a within-subjects (repeated-measures) design, every participant does every condition – for example, both congruent and incongruent Stroop trials. This controls for individual differences and needs fewer participants, but order effects such as practice and fatigue must be handled by counterbalancing or randomising trials. In a between-subjects design, different groups do different conditions, as in Loftus and Palmer’s verb groups. This avoids carry-over effects (once you have heard “smashed”, you cannot un-hear it) but needs more participants and random allocation to groups. Many studies mix the two.

Common cognitive tasks and what they measure

Much research uses standardised computer tasks, each designed to isolate one process:

Strengths and weaknesses of experiments

Experiments allow causal conclusions and are easy to replicate, but they have well-known limits. Tasks can be artificial (low ecological validity), participants are often university students in Western countries, and people may guess the hypothesis and change their behaviour (demand characteristics). Most importantly, behaviour alone often cannot distinguish between theories that make the same predictions – which is why cognitive scientists turn to the brain and to models.

Neuropsychology and Case Studies

Some of the most influential evidence in cognitive science comes from people whose brains were damaged by injury, surgery or disease. If damage to one region impairs ability A but not B, and damage elsewhere impairs B but not A, this double dissociation suggests the two abilities depend on separate systems.

• Henry Molaison (HM). In 1953 surgeon William Scoville removed parts of both of HM’s medial temporal lobes, including much of the hippocampus, to treat severe epilepsy. Afterwards HM could no longer form new long-term memories of facts and events, yet his short-term memory was normal and he improved at skills such as mirror drawing without remembering having practised. Studies by Brenda Milner and others showed that memory has several separable systems.
• Phineas Gage. In 1848 an iron rod passed through the front of Gage’s skull in a railway construction accident. His physician John Harlow reported marked changes in his personality and self-control, an early hint that the frontal lobes support planning and social behaviour. Historians note that later retellings exaggerated the changes, and that Gage seems to have recovered considerably.
• Broca’s and Wernicke’s patients. Paul Broca (1861) and Carl Wernicke (1874) linked damage in different left-hemisphere regions to different language problems, laying the foundations of neurolinguistics.

Case studies are rich but hard to generalise: brain damage rarely respects functional boundaries, and the brain reorganises after injury. Modern studies use larger groups and lesion-mapping statistics.

Neuroimaging: EEG, MEG, fMRI, PET and fNIRS

Imaging methods show what the brain is doing while people perform cognitive tasks. They differ mainly in how precisely they capture when and where activity happens.

• EEG and event-related potentials (ERPs). Electrodes on the scalp record electrical activity with millisecond precision. Averaging over many trials produces ERPs such as the N400, a response that is larger when a word does not fit its sentence context. EEG is cheap and safe for infants, but it is hard to tell exactly where signals originate.
• MEG (magnetoencephalography). Measures the tiny magnetic fields produced by neural currents. It has EEG’s timing with better localisation, but scanners are expensive and rare.
• fMRI (functional magnetic resonance imaging). Measures the blood-oxygen-level-dependent (BOLD) signal, which rises a few seconds after local neural activity. It maps the whole brain at millimetre resolution and is the dominant method in cognitive neuroscience, but it is slow and correlational: a region being active does not prove it is necessary.
• PET (positron emission tomography). Uses radioactive tracers to measure metabolism or the binding of neurotransmitters such as dopamine. It was the main imaging method for cognition in the 1980s and is still used for neurochemistry and dementia research.
• fNIRS (functional near-infrared spectroscopy). Shines near-infrared light through the scalp to measure blood oxygenation in the outer cortex. It is portable and tolerant of movement, making it useful with babies and in natural settings.

Brain Stimulation: TMS and tDCS

Imaging is correlational; stimulation can test causes. Transcranial magnetic stimulation (TMS), introduced by Anthony Barker and colleagues in 1985, uses a magnetic pulse to briefly disrupt or excite a small area of cortex. If disrupting a region slows a task, that region is probably involved in it – a “virtual lesion” in a healthy volunteer. Transcranial direct current stimulation (tDCS) passes a weak current across the scalp to nudge excitability up or down. Its effects on cognition are generally small and have been inconsistent across studies, so results should be read cautiously. In animal research, optogenetics and chemogenetics allow precise control of specific neuron types, and single-cell recording reveals how individual neurons code information.

Computational Modelling

A computational model turns a verbal theory into a program that can be run. If the model produces the same pattern of reaction times, errors or learning curves as people do, the theory gains support; if it fails, the theory must change. Major families include symbolic models and cognitive architectures such as ACT-R and Soar; connectionist (neural network) models, which learn from examples; Bayesian models, which treat cognition as probabilistic inference (an idea developed in predictive processing); reinforcement-learning models of reward-based learning; and evidence-accumulation models such as the drift-diffusion model of decisions. Today researchers also compare human behaviour and brain activity with deep neural networks and large language models. Modelling is especially important for problem solving and reasoning, where it began with Newell and Simon’s programs.

Eye Tracking

Eye trackers record where people look, for how long and in what order. In reading research, fixation durations reveal which words are hard to process. In the visual world paradigm (Tanenhaus and colleagues, 1995), listeners look at objects on a screen while hearing a sentence, and their eye movements show how quickly they interpret each word. Eye tracking is also widely used in usability testing and in studies of infants, who cannot press buttons but reliably look longer at surprising events.

Online, Corpus and Large-Scale Studies

Since the 2010s many experiments run online through platforms such as Prolific or Amazon Mechanical Turk, using software such as jsPsych, PsychoPy or Gorilla. Online testing provides larger and more diverse samples quickly, and well-established effects such as Stroop and flanker interference replicate well online, though researchers must check attention and data quality carefully. Language researchers analyse corpora – large collections of text or transcribed speech, such as the CHILDES database of child language – and “megastudies” collect responses to thousands of words from hundreds of people. See language processing for examples.

Replication and Open Science

In 2015 the Open Science Collaboration attempted to replicate 100 published psychology studies and obtained statistically significant results in only about 36% of them. Cognitive psychology studies held up better than many social psychology studies, and core effects such as Stroop interference, the flanker effect and the testing effect are highly reliable, but some well-known findings did not replicate. The field has responded with preregistration (stating hypotheses and analyses before collecting data), registered reports, larger samples, multi-lab replication projects and open sharing of data and code. When you read about a striking new finding, it is worth asking whether it has been replicated.

Choosing the Right Method

No single method answers every question. A typical research programme might start with a behavioural effect, build a model that explains it, use fMRI or EEG to see which brain systems are involved, and use TMS or patient studies to test whether those systems are necessary. In Marr’s terms, behaviour and models speak to what is computed and how, while neuroscience methods speak to how it is implemented. For key terms used on this page, see the glossary; for how these skills translate into jobs, see careers in cognitive science.