The scientific method is a set of disciplined ways to ask questions, collect evidence, test explanations and correct mistakes. Observation, hypotheses, experiments, statistics and modelling all matter, but no single mechanical sequence fits every branch of science. The standard is evidence that can be examined, challenged and improved.
What is the scientific method for students and researchers? How do control groups, measurement errors and replication work? This comprehensive Gold Standard guide gives clear answers, 25 subject chapters, four worked scientific cases, key research safeguards and links throughout the eduKate ecosystem.
Five routes into scientific reasoning
Question, hypothesis and observation · Chapters 1–5 · Start reading →
Design trustworthy research · Chapters 6–10 · Start reading →
Measure, analyse and model · Chapters 11–17 · Start reading →
Worked scientific reasoning cases · Chapters 18–21 · Start reading →
Ethics, error and knowledge transfer · Chapters 22–25 · Start reading →
Full index · Connected subjects · The Gold Standard master
Chapter-by-chapter contents
Question, hypothesis and observation
- 1. The scientific method is a discipline of checking claims
- 2. Begin with a question that can actually be answered
- 3. Observation is information, not an untouched window onto reality
- 4. Construct hypotheses with observable consequences
- 5. Prediction and explanation serve different roles
Design trustworthy research
- 6. Set the measurement definition before data collection
- 7. Controls and comparison groups protect causal questions
- 8. Randomisation, blinding and allocation
- 9. Observational research requires special care
- 10. Sampling frames, missing data and representativeness
Measure, analyse and model
- 11. Measurement error and uncertainty budgets
- 12. Units, scales and dimensional checks
- 13. Collect data with a record that another person can audit
- 14. Analysis plans, statistical models and assumptions
- 15. Separate exploratory discovery from planned confirmation
- 16. Build models, then try to break them
- 17. Triangulation and independent evidence
Worked scientific reasoning cases
- 18. Worked example 1: one variable changes, but the baseline does too
- 19. Worked example 2: the precise thermometer with a biased reading
- 20. Worked example 3: relative and absolute event differences
- 21. Worked example 4: repeat a calculation or repeat a study?
Ethics, error and knowledge transfer
- 22. How results can fail—and why that is informative
- 23. Ethics, safety, consent and responsible research
- 24. From school investigation to professional science
- 25. Questions, specialist directory and the continuing method
Four things an investigation must reveal
| Question | Good evidence practice | Common error |
|---|---|---|
| What is being claimed? | Define the tested quantity and scope | Shifting the claim after results |
| What was observed? | Report units, sources and procedures | Confusing sensor output with the target quantity |
| What explains it? | Compare controls and alternatives | Treating association as proof of causation |
| How robust is it? | Quantify uncertainty, reproduce and test again | Reporting one favourable result as universal |
1. The scientific method is a discipline of checking claims
Scientific method means using systematic observation, reasoning, experiments, modelling and criticism to build explanations that can be corrected by evidence. It is better understood as a family of rigorous practices than as one magical six-step worksheet.
A historian comparing documents, an astronomer studying stars and a biologist running a controlled experiment may not use identical procedures. Their common commitment is to make questions, assumptions, observations and reasoning available for challenge.
Why there is no universal laboratory recipe
A field astronomer cannot randomly assign galaxies to treatment groups, and a geologist cannot restart an earthquake to hold one variable fixed. They can still make rigorous claims by comparing evidence, using mathematical models, measuring independent signals and testing predictions. A laboratory experiment is a powerful method, not the definition of all science.
Teach learners to ask which form of evidence is possible for the question. If deliberate manipulation is impossible, what observations would distinguish alternative explanations? If the result is only a model prediction, what independent measurements might challenge it? This habit preserves scientific standards without pretending all knowledge must arrive through the same classroom worksheet.
2. Begin with a question that can actually be answered
A useful investigation identifies a phenomenon, a population or system, the quantity of interest and a feasible source of evidence. ‘Why does everything change?’ is a stimulating conversation starter but not yet a tractable measurement plan.
Move toward a precise question such as: under stated classroom conditions, does a change in revision interval alter recall at a later test? Then decide what would count as a meaningful difference before seeing the results.
Make the research question small enough to be useful
Suppose the starting question is ‘Does background music help studying?’ Before collecting data, define the kind of music, the task, the learners, the duration and what improvement means. Does the study ask about reading comprehension, memorisation, problem-solving accuracy or persistence? Those outcomes might move in different directions.
A tractable version could compare immediate recall scores for a specified reading task under two sound conditions, using appropriate allocation and comparable material. That does not answer every question about music and learning. Its strength is that readers can see precisely what was tested and where the conclusion stops.
3. Observation is information, not an untouched window onto reality
A telescope records light through an instrument. A thermometer responds according to a calibration. A survey records answers shaped by wording and context. Even seemingly simple observations reflect choices of method and interpretation.
Write down where the observation came from, which unit it represents and which transformation was applied. ‘The meter displayed 7’ is not yet evidence of seven grams, centimetres or units of exposure.
Instruments carry assumptions into the data
Imagine a plant-growth investigation using a phone photograph each morning. The apparent height of the plant could change because the camera angle, distance or reference ruler moved. The recorded image is real, but its measurement interpretation still depends on the setup. A better design holds camera placement steady and measures against a reference scale.
Students can test this principle by photographing the same object from different angles and comparing apparent dimensions. Their conclusion is not that photographs are unreliable in general, but that instrument position is part of the method. Precise observation requires a chain from the physical quantity to the recorded number.
4. Construct hypotheses with observable consequences
A hypothesis proposes a relation or explanation that can be examined. It becomes informative when it produces clear expectations that differ from plausible alternatives. Hypotheses may be refined or rejected as new evidence appears.
An untestable assertion that adapts to every possible result is not a useful investigative claim. Yet a prediction turning out wrong is often valuable; it may expose a hidden assumption, a measurement flaw or an incomplete theory.
Design predictions that could genuinely surprise you
A hypothesis becomes informative when it distinguishes between competing outcomes. If someone claims a study technique works ‘whenever the learner is ready’, every failed observation can be dismissed after the fact. A clearer claim predicts measurable differences for specified tasks and conditions, and says in advance what result would count against it.
Scientists may still update their hypotheses as they learn. The important distinction is whether a proposed explanation was tested with evidence not already used to construct it. New hypotheses developed after looking at data deserve new tests; treating them as pre-existing predictions makes evidence sound stronger than it is.
5. Prediction and explanation serve different roles
A predictive model can forecast an outcome without revealing why it occurs. A causal explanation aims to describe mechanisms and the conditions in which they operate. Science uses both, but they are not interchangeable.
For example, a weather model can be evaluated by forecast skill while researchers investigate the physical mechanisms separately. A high-accuracy classifier alone does not establish causation.
Prediction can succeed for the wrong reason
A model may estimate test scores by using students’ previous results. It could perform well while revealing little about which teaching methods cause improvement. Similarly, a predictive system may identify a disease risk marker that is not itself responsible for the disease.
To move from prediction to explanation, ask what would happen if the proposed causal factor were changed while the relevant alternatives were addressed. That often demands a different study design. Good scientific communication celebrates reliable prediction where useful without quietly renaming it an established mechanism.
6. Set the measurement definition before data collection
An operational definition states how a concept is measured: what counts as a success, how long the test lasts, how the instrument is calibrated, and which cases will be included or excluded.
A vague outcome such as ‘became better’ allows almost any improvement to count after the fact. A clear measure might be the percentage of independently solved algebra problems on a new assessment at a defined time.
When two people score the same answer differently
An outcome such as ‘quality of explanation’ needs a scoring rubric that another evaluator can apply. Without examples and criteria, one reader may reward technical vocabulary while another rewards sound reasoning. The apparent effect of an intervention can then be dominated by how outcomes were graded.
Build an operational definition with observable elements: correct relation, appropriate units, justified inference and explicit limit, for example. Try the rubric on a few sample answers before the main study. Compare assessors, discuss ambiguous cases and record revisions. Agreement alone does not prove that the rubric measures every aspect of scientific understanding, but it identifies one important source of variability.
7. Controls and comparison groups protect causal questions
Experiments seek to isolate the effect of a change by comparing conditions. The comparison group shows what might happen without that specific intervention, while controlled variables and consistent procedures reduce alternative explanations.
A perfect control is not always feasible or ethical. Researchers should explain which differences remain and how those limits affect causal claims rather than pretending the comparison is flawless.
The control group’s real job
If all students improve between September and November, a programme may deserve credit—or the improvement may reflect normal coursework, extra practice, maturity or an easier test. A comparison group attempts to estimate what would have happened without the specific programme.
The comparison must be appropriate to the question. If the groups differ greatly at the beginning, comparing their final means alone can mislead. If one group receives more contact time, the study may confound the intended method with the amount of attention. Controls are not decorative; they make the alternative explanation explicit.
8. Randomisation, blinding and allocation
Random assignment can reduce systematic differences between study groups when implemented properly. Blinding, when feasible, can reduce some expectation and measurement effects. Neither automatically fixes poor measurements, loss to follow-up or a tiny sample.
The question is always whether the research design makes the counterfactual credible: what would likely have happened to the same participants without the intervention? That cannot be observed directly, so the design supplies a reasoned approximation.
Allocation concealment and blinding solve different problems
Random assignment aims to make group allocation independent of participant characteristics, within the design’s assumptions. Allocation concealment aims to prevent advance knowledge of assignments from influencing enrolment. Blinding, where feasible, limits knowledge of assignments during treatment or measurement. These are related safeguards but not synonyms.
In an education study students and teachers may know which lessons they receive, so full blinding may be impractical. Independent scoring of unfamiliar assessment responses may still reduce expectation effects. Researchers should state what was possible, which risks remain and what design choices were made to address them.
9. Observational research requires special care
Much important science is observational: astronomy, ecology, epidemiology and historical climate research cannot always manipulate their subjects. Researchers analyse naturally occurring variation, compare cases and test competing explanations.
Observational evidence can be powerful, but associations can reflect confounding, selection or reverse causation. State the design and the identification assumptions before describing a result as a treatment effect.
An observational study can test competing mechanisms
Suppose city areas with more trees are also cooler. Trees may contribute shade and evaporation, but the areas may differ in building density, land cover, wind or measurement time. An observational researcher can collect these additional variables, compare suitable locations and examine the predicted physical mechanisms.
No single statistical adjustment guarantees causation, especially if relevant factors remain unmeasured. Multiple lines of evidence can still make the explanation more persuasive. Good reporting says how comparisons were constructed and which untested alternatives might still account for the observed pattern.
10. Sampling frames, missing data and representativeness
A sample represents the set of observations actually obtained, not automatically all people, places or events of interest. Eligibility rules, voluntary participation, nonresponse and incomplete records may distort estimates.
Data that are missing because of the outcome being measured can be particularly troublesome. A larger dataset does not guarantee less bias when the process that selected the observations is systematically unrepresentative.
Why a bigger sample sometimes preserves a bad answer
A poll of 100,000 self-selected website visitors can estimate that visitor group’s responses very precisely. It may nonetheless be a poor guide to people who never saw the website or chose not to respond. Larger numbers do not automatically cure selection bias.
Describe the target population and how a person could enter the dataset. Then examine nonresponse, eligibility and missing observations. If the sample systematically excludes a group most affected by the issue, the resulting precision may be deceptive. The correct repair usually concerns study design or careful qualification, not adding decimals.
11. Measurement error and uncertainty budgets
Every measurement has limits. Precision concerns how closely repeated measurements agree; accuracy concerns closeness to a relevant reference value. A device may be consistently precise and still biased.
Record calibration, environmental conditions, resolution, drift, operator procedure and uncertainty. Separate random variation from systematic error: averaging many readings may reduce random noise, but it does not necessarily repair a miscalibrated instrument.
Random error and systematic bias behave differently
Repeated measurements of a stable object may fluctuate because of ordinary noise. Their mean can sometimes reduce that random variation. If every measurement is shifted by the same calibration error, averaging more of them may produce a very precise but still biased estimate.
Ask learners to separate spread from offset in a graph. A cluster of results tightly grouped around the wrong reference is precise but inaccurate. One widely scattered set centred on the correct reference may have small average bias yet poor repeatability. Correct conclusions require an uncertainty account that matches the kinds of error present.
12. Units, scales and dimensional checks
A physically meaningful calculation requires compatible units and an appropriate model. Adding metres to seconds, treating percentages as raw counts, or confusing degrees Celsius with kelvin in the wrong formula can invalidate an otherwise polished result.
Dimensional analysis is an inexpensive error detector. Before trusting a computed speed, check that its units reduce to distance per time. State whether a numerical value is observed, derived, rounded or illustrative.
A units checklist that catches expensive mistakes
Before calculating a speed, check that distance and time use compatible units. Before calculating a density, check whether mass and volume refer to the same sample. Before adding probabilities, consider whether the events overlap. In scientific work, simple dimension and definition checks often catch errors sooner than complex analysis.
Try a changed-context case: a satellite speed is given in kilometres per second, while a planned manoeuvre distance is in metres. Converting one unit consistently is necessary before division. Even if a calculator returns a neat number, a result with nonsensical units is a warning that the equation, inputs or both need revision.
13. Collect data with a record that another person can audit
Keep timestamps, equipment or software versions, procedures, sample identifiers and reasons for exclusions. Data tables should preserve raw measurements before transformations and corrections.
A good laboratory notebook includes mistakes and unexpected results as well as successes. Rewriting the record to make the experiment look tidy sacrifices the evidence needed to identify what happened.
Why a notebook needs unsuccessful attempts
An experiment that repeatedly fails to stabilise at a target temperature is not useful only when the apparatus finally works. Its earlier instability may reveal instrument response, a hidden external heat source or problems with sample handling. Removing failed attempts from the record erases the trail that helps the next person reproduce a trustworthy result.
Record what was planned, what happened, which instrument settings changed and why observations were excluded. The notebook is not a performance review. It is a record of how knowledge was built. Students who learn to preserve mistakes are better prepared to diagnose them.
14. Analysis plans, statistical models and assumptions
Statistical methods turn observations into estimates, uncertainty summaries and tests of hypotheses. Their meaning depends on model choice, sample design, independence, distributions and analytic decisions.
A small p-value is not the probability that the scientific hypothesis is true. A confidence interval is not simply the range containing most raw measurements. Interpreting results demands a correct account of what the procedure actually estimates.
Statistical significance is not a synonym for importance
An effect can produce a small p-value in a large study while being too small to matter in practice. Conversely, a small study may estimate a potentially important effect imprecisely. Decisions should consider the estimated magnitude, uncertainty, study design and consequences, not one threshold.
If a report gives p = 0.03, ask which null model and assumptions produced it, whether the analysis was planned, how many comparisons were tried and what the actual effect size was. A p-value is not the chance that the whole scientific hypothesis is false. Precision in language is as important as precision in arithmetic.
15. Separate exploratory discovery from planned confirmation
Exploratory analysis searches for patterns and develops new questions. Confirmatory analysis tests questions and rules defined beforehand. Both are legitimate when clearly labelled.
If researchers try many outcomes and report only the most favourable result as if it had been predicted, readers cannot assess the rate of false positives fairly. Transparent hypotheses, preregistration when appropriate and external tests can strengthen subsequent inference.
A discovery is not the same as its independent confirmation
Data exploration is valuable because unfamiliar patterns can suggest new science. But when a researcher examines dozens of candidate relationships and selects the most striking one, it should usually be presented as an exploratory finding. The same dataset is not fully independent evidence for both discovering and confirming that pattern.
A new sample, suitable validation set or an independently planned test can help. A preregistered plan is one useful safeguard, though not every form of research fits the same formal process. The principle is transparent separation between questions developed before observing an outcome and hypotheses invented after seeing it.
16. Build models, then try to break them
A model simplifies a complicated system into variables, relationships and boundary conditions. Its value is measured by how well it serves a defined purpose, not by how elegant its equations or code appear.
Check input data, initial conditions, sensitivity to parameters and performance outside the training or calibration sample. A model validated for one environment may fail in another, even when the calculations are internally consistent.
Verification asks whether the model was solved correctly
Verification and validation address different questions. In computational work, verification checks whether equations and numerical methods were implemented and solved as intended. Validation checks whether the model represents the real phenomenon accurately enough for a specified use. A beautifully verified computer programme can still model the wrong physical process.
Test an implementation against cases with known answers, check numerical convergence where relevant, and examine sensitivity to input choices. Then seek independent empirical observations. If the model is used outside its original setting, repeat the validation question: sufficient for what decision, at what conditions and with what uncertainty?
17. Triangulation and independent evidence
Confidence often increases when evidence from different methods points toward compatible explanations. A microscopic observation, a chemical measurement and a functional experiment may constrain different parts of one proposed mechanism.
But apparently independent measurements can share one calibration error or dataset. A good triangulation argument explains the distinct assumptions rather than simply counting the number of agreeing charts.
Independent measurements must be genuinely independent
Suppose three temperature sensors all report a warming trend. If all were calibrated against the same biased reference, agreement may reflect a shared error. Triangulation becomes persuasive when methods have different limitations and still support compatible explanations.
Compare measurement principles, locations, calibrations and data processing. One method may reveal a rapid change but not its chemical cause, while another offers chemical identification but only at a few times. Together they can narrow the possibilities. Counting three graphs is less informative than explaining the independent evidence each supplies.
18. Worked example 1: one variable changes, but the baseline does too
Invented classroom trial: students complete 12 problems before a tutorial and 15 afterward. The raw increase is 3 questions, or 25% relative to the original 12. But the later test may be easier, the learners more familiar with the format, or the questions not comparable.
A defensible improvement claim needs an appropriate comparison and a clear scoring method. Arithmetic proves that the score rose in this exercise; it does not by itself prove which teaching method caused the change.
19. Worked example 2: the precise thermometer with a biased reading
Invented calibration check: a thermometer reports 22.4°C, 22.5°C and 22.4°C for a reference condition that is independently established as 23.0°C. The mean is 22.433…°C, close to 22.43°C. The readings are tightly clustered, but the average is about 0.57°C low.
More repeated readings could make the mean more precise without removing the offset. Calibrate or correct the instrument using a valid procedure, and report the uncertainty of the correction.
20. Worked example 3: relative and absolute event differences
Invented event counts: in two equal groups of 1,000 people, one has 40 observed events and the other 20. Event rates are 4% and 2%, an absolute difference of 2 percentage points and a relative reduction of 50%.
Without allocation details, follow-up times, uncertainty intervals or subgroup context, these figures do not establish the cause. Two mathematically correct percentage descriptions can imply very different practical magnitudes.
21. Worked example 4: repeat a calculation or repeat a study?
Two distinct tasks: an independent researcher reruns a published analysis using exactly the original data and code and obtains the same 0.42 estimate. That supports computational reproducibility under the specified conditions.
A second team collects new measurements under a comparable design and estimates 0.39. That is a new-data replication attempt, whose interpretation depends on uncertainty, methods and context. The National Academies distinguish the two because they test different questions.
Reproducibility and replication as two different promises
Computational reproducibility asks whether a result can be obtained from the same data and methods under the documented conditions. Replicability asks whether comparable findings emerge when new data are obtained in a suitable new study. A programme can faithfully reproduce a biased result; a new experiment may uncover that bias.
To support reproducibility, preserve analysis scripts, parameters, transformations and justified exclusions where legally and ethically possible. To support replication, describe sampling, interventions, instruments and contextual constraints well enough for another team to design a fair test. Both forms of openness help science self-correct, but neither guarantees agreement under changed conditions.
22. How results can fail—and why that is informative
Null results, disagreement and unsuccessful replications can expose flawed assumptions or genuine differences between populations and conditions. A failed result does not automatically mean fraud or useless research.
Investigate protocol drift, instruments, sampling, coding mistakes, statistical power and context. Preserve the possibility that the original hypothesis may need revision. Scientific resilience means learning from disagreement, not hiding it.
23. Ethics, safety, consent and responsible research
Research must respect people, animals, communities and environments. Obtain necessary ethical review and consent, protect private information and avoid exposing participants or researchers to unjustified hazards.
The fact that an experiment could answer a question does not mean it is permissible. Safer designs, observational alternatives, simulation or published datasets may achieve useful learning without unacceptable risk.
24. From school investigation to professional science
For younger learners, begin with observation, one clearly changed factor, consistent measurement and an honest conclusion. Secondary and university learners can progress to controls, uncertainty, sampling, competing models and reproducibility.
The teaching goal is not to memorise the sequence ‘question–hypothesis–experiment–conclusion’. It is to recognise where a claim needs a better test. A mentor can teach from mistakes by asking students which assumption broke.
25. Questions, specialist directory and the continuing method
Does all science require a laboratory experiment? No. Some fields rely strongly on observation, models, comparison and inference. Does one failed experiment disprove an entire theory? Not necessarily; the test and auxiliary assumptions need scrutiny.
Does a statistically significant result prove importance? No. Report effect size, uncertainty and real-world consequences. How can I learn scientific method? Choose one question, state what would change your mind, measure carefully, then explain your conclusion and its limits.
The connected eduKate scientific reading directory
Reasoning and quantitative foundations
- The Gold Standard of Evidence-Based Thinking
- The Gold Standard of Analytical Thinking
- The Gold Standard of Mathematical Modelling
- The Gold Standard of Biostatistics
- The Gold Standard of Uncertainty Quantification
- The Gold Standard of Numerical Analysis
Methods in real disciplines
- The Gold Standard of Signal Processing
- The Gold Standard of Structural Biology
- The Gold Standard of Remote Sensing | NDVI Averaging
- The Gold Standard of Systems Thinking
- The Gold Standard of Bioinformatics
Educational and practical applications
- The Gold Standard of Science Communication
- The Gold Standard of School Citizen Science Data Contributions
- Gold Standard Case Study 209 | SD or Confidence Interval?
- The Gold Standard of Education | Learning Transfers into Life
Scholarly method references
- Stanford Encyclopedia of Philosophy — Scientific Method, revised 2026
- Stanford Encyclopedia of Philosophy — Theory and Observation in Science
- National Academies — Reproducibility and Replicability in Science (2019)
- National Academies — Reproducibility and Replicability definitions
- Stanford Encyclopedia of Philosophy — Experiment in Physics
Five questions before claiming a result
- Which observation would change the conclusion?
- What instrument, unit, data selection and control were used?
- Which alternative explanation has not been ruled out?
- What uncertainty and limits remain?
- Can another researcher audit the calculation and test the claim with new evidence?
Continue through The Gold Standard master, or apply the method in Science Communication.
Continue with the wider How X Works library and The Gold Standard of Education. The goal of scientific learning is to carry a method into a new context, question its assumptions and recognise when the evidence calls for a better explanation.