How Scientific Research Works | From a Question and Prior Knowledge to Method, Evidence, Criticism, Replication and Correction

Scientific research works by turning an important uncertainty into a question that can be investigated, designing a disciplined way to encounter relevant evidence, analysing that evidence without pretending it says more than it does, exposing the claim to criticism and independent checking, and allowing later evidence to strengthen, narrow, overturn or replace the answer.

That sentence is longer than the school version of the scientific method for a reason.

Real research is not usually a neat staircase marked question → hypothesis → experiment → conclusion. Sometimes a new instrument reveals a pattern nobody was looking for. Sometimes researchers begin with a theory. Sometimes they begin with a dataset, a strange observation, an engineering failure, a historical archive, a field interview, a mathematical inconsistency or a public problem. Sometimes the first study is exploratory. Sometimes the decisive evidence arrives years later from another laboratory, another population or another method.

What makes the activity scientific is not obedience to one classroom sequence. It is the harder discipline underneath: claims must remain answerable to evidence and correctable by the world.

The central question: what would we have to observe, measure, compare, calculate or independently recover before this claim deserves more confidence than its alternatives?

Quick Read

UNCERTAINTY → QUESTION → PRIOR KNOWLEDGE → ALTERNATIVE EXPLANATIONS → STUDY DESIGN → SAMPLING → MEASUREMENT → ETHICS → DATA → ANALYSIS → UNCERTAINTY → CLAIM → CRITICISM → REPRODUCIBILITY → REPLICATION → SYNTHESIS → CORRECTION → BETTER QUESTION

The chain loops. It branches. Parts can occur in a different order. Some research has no experiment. Some has no formal null hypothesis. Some produces descriptions rather than causal estimates. Some builds instruments, maps, taxonomies, simulations or theories. But strong research keeps the connection between the question, the method, the evidence and the strength of the conclusion visible enough that other people can inspect it.

1. Research Begins Before the First Experiment

Imagine a student asking, “Does background music improve concentration?”

It sounds like a research question. But almost every important word is still unstable.

• What counts as background music?
• What kind of concentration?
• For which people?
• Doing what task?
• For how long?
• Compared with silence, white noise or another condition?
• Does “improve” mean speed, accuracy, memory, persistence or subjective comfort?
• Could music help one task and interfere with another?

This is why good research often begins with decompression. A broad curiosity has to be unpacked into a form that can meet evidence.

The first skill is therefore not experimentation. It is learning to ask a question whose parts are clear enough that a result could actually change what we believe.

2. There Is No Single Scientific Method

The National Academies describes science as a mode of inquiry built around questions about the world, evidence, reasoning, communication and communal checking. It also makes an important point that school diagrams can accidentally hide: there is no fixed sequence of steps that every scientist always follows.

Astronomers cannot move a galaxy into a control group. Geologists cannot rerun the formation of a mountain range. Epidemiologists may study naturally occurring differences. Ecologists may combine field observation with experiments. Physicists may infer invisible entities from measurable effects. Computer scientists may build algorithms and benchmark them. Social scientists may use surveys, interviews, field experiments, administrative data and natural experiments. Historians use surviving records and source criticism. Mathematicians may prove rather than empirically test.

The surface methods differ because the world presents different kinds of problems.

The deeper rule is more stable:

A claim should be no stronger than the evidence-producing route that connects it to reality.

3. The Question Determines the Research Job

“Research” is too broad to be one job. Different questions require different kinds of evidence.

Question type	Typical research job	Example
Description	Measure what exists, how often, where or in what form.	How common is a behaviour in a defined population?
Comparison	Estimate differences between conditions, groups, places or times.	Do two teaching approaches produce different outcomes?
Causation	Distinguish an effect from alternative explanations and confounding.	Did the intervention cause the improvement?
Mechanism	Explain the process that connects cause and effect.	Through what pathway does sleep influence memory?
Prediction	Estimate future or unobserved outcomes and validate performance.	Can these variables forecast equipment failure?
Exploration	Find patterns, categories, anomalies or candidate explanations.	What structures appear in a new dataset?
Evaluation	Judge whether a programme, policy or system works for its intended purpose.	Did a transport change reduce journey time without shifting harm elsewhere?
Interpretation	Understand meaning, experience, context or social process.	How do participants experience a new policy?

Trying to answer a causal question with purely descriptive evidence is one of the oldest research errors. So is demanding a randomised experiment for a question that cannot ethically or practically be randomised.

4. Good Research Knows What Is Already Known

Before collecting new data, researchers usually need to understand the existing evidence landscape.

This is not ceremonial citation. It changes the design.

• Has the question already been answered well?
• Where do studies disagree?
• Which populations have been neglected?
• Which measurements are reliable?
• Which methods repeatedly fail?
• What effect sizes are plausible?
• What mechanisms have been proposed?
• What assumptions are hidden inside the standard explanation?

A literature review is therefore part map, part memory and part error detector. It helps the new study start from humanity’s accumulated knowledge rather than from zero.

This is also why citation, libraries, archives, search and indexing belong inside the research system. Discovery infrastructure affects what researchers can know about what has already been done.

5. Research Questions Need Boundaries

Suppose we ask, “Does exercise improve health?”

That question is too compressed.

Which exercise? What intensity? How often? Which health outcome? In what age group? Over what period? Compared with what? Does the person already have a condition that changes the effect?

Research requires a system boundary. The boundary does not claim the rest of reality is irrelevant. It says: for this investigation, these are the units, conditions, variables, timescales and outcomes we will make visible.

The moment we draw that boundary, we should also ask what it excludes.

Boundary check: what could matter to the answer but fail to enter the study?

6. Hypotheses Are Useful — but Research Is Larger Than Hypothesis Testing

A hypothesis is a proposed explanation or relationship that can be confronted with evidence. It can turn a vague idea into expected observations. If one explanation is true, what should we see? If a rival explanation is true, what should be different?

But not every valuable study begins with a formal hypothesis. Exploratory research can reveal previously unknown patterns. Descriptive work can establish baselines. Instrument development can make later tests possible. Qualitative research can uncover mechanisms or categories that were not adequately specified in advance.

So a world-class research culture distinguishes two legitimate jobs:

• Exploration: what might be here?
• Confirmation: does this pre-specified claim survive a fair test?

The danger comes when exploratory choices are made after seeing the data but reported as though they were planned all along. That makes the evidence look cleaner and more decisive than the actual route justified.

For the deeper mechanism, see How Hypotheses Work.

7. The Study Design Is an Evidence Machine

A study design is not formatting. It is the architecture that determines what comparisons the evidence can support.

If you want to estimate a treatment effect, the design must help distinguish the treatment from what would have happened anyway. If you want prevalence, the sampling route must represent the target population appropriately. If you want to understand a process, the observations must be rich enough to discriminate among mechanisms. If you want prediction, the model must be tested on data that genuinely challenge generalisation.

Design family	What it can help reveal	Typical weakness to watch
Randomised experiment	Causal effects under defined conditions	Artificial settings, attrition, implementation failure, limited generalisability
Observational cohort	Associations and temporal patterns	Confounding, selection, measurement differences
Case-control study	Factors associated with a relatively rare outcome	Selection and recall bias
Cross-sectional survey	Snapshot of a population or relationship	Sampling bias and weak temporal inference
Natural experiment	Causal clues from external variation	Assumptions about comparability
Qualitative study	Meaning, context, process, lived experience	Overgeneralising beyond the evidence base
Simulation/model	Behaviour under explicit assumptions	Model misspecification and false realism
Systematic review/meta-analysis	Synthesis across studies	Bad inputs, heterogeneity, publication bias

No design is simply “best”. The important question is: best for which inference?

8. Sampling Decides Which Part of Reality Gets a Vote

Most research cannot observe every possible person, object, event, molecule, school, city, organism or time point. It works with samples.

A sample is therefore not a clerical subset. It is a gate between a large target and the evidence that enters the study.

A large sample can still be badly biased. Ten thousand volunteers from the wrong sampling frame do not automatically represent the people the claim is about. A small, carefully selected sample can sometimes be much more informative for a narrow question than an enormous convenience sample.

Sampling therefore asks who or what had a chance to enter, who did not, who refused, who disappeared during follow-up, and whether those missing pathways could change the result.

See How Sampling Works for the full treatment.

9. Measurement Determines What the Study Actually Sees

Researchers often discuss concepts that cannot be observed directly.

Stress. Learning. Poverty. Biodiversity. Air quality. Intelligence. Trust. Disease severity. Political participation. System reliability.

To study them, researchers need operational definitions, instruments, coding rules, sensors, rubrics, assays or other procedures that turn part of the world into recorded evidence.

The measurement is not the thing itself. A test score is not the whole learner. A questionnaire item is not the whole emotion. A satellite pixel is not the entire forest. A biomarker may be informative without being identical to a disease state.

This is where How Measurement Works becomes foundational. Calibration, validity, precision, uncertainty, missingness and proxies can determine whether a beautifully analysed dataset is actually about the question researchers think they asked.

10. Controls, Comparisons and Counterfactuals

Causal questions contain a hidden difficulty.

If a student receives a new teaching method and improves, we observe what happened with the intervention. We do not simultaneously observe what would have happened to that same student at the same time without it.

That missing alternative is the counterfactual.

Research designs create approximations to the missing comparison through control groups, randomisation, matching, natural experiments, longitudinal comparisons, statistical adjustment and other strategies. None literally opens a second universe. They create more or less defensible bridges to the unobserved alternative.

This is why counterfactual reasoning sits beneath much of causal research.

11. Randomisation Solves a Particular Problem

Randomisation is often misunderstood as a magic quality stamp.

In an experiment, random assignment can help balance both known and unknown differences between treatment groups, making the treatment contrast more credible. But random assignment is not the same as random sampling. The first is mainly about causal comparability inside the study. The second is about how the sample relates to a wider population.

A perfectly randomised trial in a narrow group can estimate an internally credible effect and still leave difficult questions about whether the result travels to different ages, countries, environments or delivery systems.

Research becomes stronger when it stops treating “randomised” as the end of thought and asks what randomisation did—and did not—solve.

12. Blinding Reduces Information Pathways That Can Distort Results

When participants, researchers, assessors or analysts know which condition a case belongs to, that knowledge can sometimes influence behaviour, measurement or judgement.

Blinding attempts to block some of those pathways.

It is not always possible. A teacher usually knows which teaching method they are using. A surgeon knows which operation was performed. A policy change may be visible to everyone. When blinding cannot be used, researchers have to think harder about alternative controls, objective outcomes, independent assessment and sensitivity analysis.

13. Ethics Is Part of Method, Not Decoration

A study does not become scientifically acceptable merely because its design could produce informative data.

Research involving people, animals, communities, sensitive records or environmental intervention carries obligations that constrain what may be done. Informed consent, privacy, risk minimisation, fair selection, independent ethics review and appropriate care for vulnerable groups are not external obstacles to science. They help define what responsible science is allowed to become.

This matters because the desire for a clean answer can create pressure to treat the subject as an instrument for the study. Ethical research reverses that priority. The person or living system is not raw material for a result.

14. Data Is a Recorded Representation of the Study, Not the Study Itself

Once evidence is recorded, the research problem changes shape.

Now there are files, labels, timestamps, units, missing values, transformations, exclusion decisions, codebooks and versions. A later analyst may never see the original field site, participant, specimen or instrument. They see a representation.

That representation needs provenance.

• Where did the data come from?
• Who collected it?
• Under what protocol?
• Which records were excluded?
• Which transformations were applied?
• How were missing values handled?
• Which version of the dataset produced the published result?

How Data Works explores this layer more fully. The key idea is simple: every dataset is a compressed historical trace of how the study encountered the world.

15. Analysis Is Where Evidence Meets a Model of What the Evidence Means

Analysis can feel objective because software returns exact numbers.

But before the software runs, people have already made choices.

• Which outcome matters?
• Which observations count?
• Which covariates enter the model?
• Which functional form is assumed?
• How are outliers treated?
• How is missingness handled?
• Which subgroup comparisons are tested?
• Which uncertainty model is appropriate?

The calculation may be reproducible while the analytic decision remains debatable. That is why transparent research reports not just the final number but enough of the path to reconstruct how the number was produced.

16. Estimation Often Matters More Than a Binary Answer

Many readers learn statistics through “significant” versus “not significant”. That can make research feel like a courtroom in which a p-value declares a hypothesis alive or dead.

Real decisions often need more.

• How large is the estimated effect?
• How uncertain is that estimate?
• Is the effect practically important?
• How sensitive is it to reasonable analytic choices?
• Does it persist across subgroups or settings?
• How does it compare with costs, risks or alternatives?

Statistical significance cannot answer all of those questions. A tiny effect can become statistically detectable in a very large sample. A meaningful effect can remain uncertain in a small study. Research therefore needs effect sizes, intervals, model checks and domain judgement—not a single ritual threshold.

17. Uncertainty Is Information

Scientific writing sometimes sounds weakest when it is actually strongest.

“The evidence suggests…”

“Within this population…”

“Under these assumptions…”

“The estimate is compatible with…”

Those phrases are not necessarily evasions. They can be signs that the author understands the boundary between observation and certainty.

Uncertainty is part of the research result. Sampling variation, measurement error, model uncertainty, unmeasured confounding and unknown external conditions all place limits on what the study can claim.

18. A Null Result Is Not the Same as “Nothing Happened”

Suppose a study does not find clear evidence of a difference.

Several realities are possible.

• There may truly be little or no effect.
• The study may have been too small or noisy to resolve the effect.
• The treatment may have been poorly implemented.
• The outcome measure may have been insensitive.
• The effect may occur only in a subgroup or context not captured by the design.
• The underlying hypothesis may have been wrong.

A good conclusion therefore distinguishes absence of evidence from sufficiently precise evidence of little or no effect. That distinction can change policy, medicine, engineering and education.

19. Exploratory and Confirmatory Research Need Different Honesty

Exploration is essential. It is how researchers notice surprising patterns and generate better questions.

But if one dataset is searched through hundreds of possible relationships, some apparently striking patterns will occur by chance. A pattern discovered after extensive searching should not be described as though it were an independently pre-specified test.

The repair is not to ban exploration. It is to label the research state correctly.

Discover here. Confirm elsewhere.

Preregistration, registered reports, held-out datasets and independent replications are different mechanisms for preserving that boundary where it matters.

20. Reproducibility Asks Whether the Analytic Route Can Be Recovered

Modern research frequently depends on computation. A published table may sit at the end of a long chain of code, packages, data cleaning, parameter choices and versioned files.

The National Academies uses a useful distinction: reproducibility can refer to obtaining consistent results using the same data, computational steps, methods and code, while replicability asks whether new data collected for the same scientific question lead to consistent conclusions.

Reproducibility therefore tests whether the reported analytic result can be reconstructed from the evidence package. It catches transcription errors, missing steps, undocumented transformations and computational failures that may otherwise hide behind a polished paper.

21. Replication Asks Whether the Result Survives New Contact With the World

A result can be computationally reproducible and still fail to appear in a new study.

That does not automatically prove misconduct or incompetence. Context may differ. Measurement may be more precise. The original effect may have been overestimated. The new study may itself be flawed. The phenomenon may be genuinely unstable. Hidden moderators may exist.

Replication is powerful because it forces a claim to leave the protected environment of its original dataset and meet another slice of reality.

Science becomes more trustworthy when the community asks not “Was this paper impressive?” but “What happens when competent people try again?”

22. Peer Review Is a Filter, Not a Truth Machine

Before publication, many research papers are examined by editors and external reviewers with relevant expertise. Reviewers may challenge methods, statistics, interpretation, missing literature, ethics, controls or overclaiming. Authors may revise. Editors decide whether the work meets the journal’s standards.

This can improve a paper enormously.

But peer review does not certify truth. Reviewers work with limited time and information. Different experts can disagree. Errors can survive. Fraud can evade review. Important work can be rejected. Weak work can be accepted.

Peer review is best understood as one checkpoint in a longer correction system. See How Peer Review Works.

23. Publication Is Not the Finish Line

A paper can be published and later criticised, corrected, re-analysed, replicated, contradicted, withdrawn or absorbed into a stronger synthesis.

This is not evidence that science is broken. The ability to change the record when better evidence appears is one of the mechanisms by which science avoids becoming a museum of permanently frozen conclusions.

The important distinction is between error and uncorrectable error. Every human knowledge system will make mistakes. A scientific culture earns trust when it makes mistakes discoverable and correction consequential.

24. One Study Rarely Owns the Whole Answer

News headlines often compress research into a dramatic form:

Scientists prove X.

Most serious research deserves a more careful frame.

One study usually contributes a weighted piece of evidence. Confidence grows when different methods, datasets, researchers, populations and disciplines converge on compatible explanations while serious alternatives lose explanatory power.

This is why systematic reviews, meta-analyses, evidence syntheses and consensus assessments matter. They attempt to reason across a body of work rather than letting the most memorable paper dominate the map.

25. Triangulation Is Powerful Because Different Methods Fail Differently

Suppose three research methods have different vulnerabilities.

A survey can reach many people but may depend on self-report. A behavioural task can directly observe performance but may be artificial. Administrative data can reflect real-world behaviour but contain categories created for operational rather than research purposes.

If all three point in the same direction, the convergence can be more informative than simply repeating one method many times. The strength comes not from counting studies but from understanding whether their error structures are independent enough to challenge one another.

26. Generalisability Is a Separate Question

Even an internally strong result lives somewhere.

Which population? Which school system? Which climate? Which species? Which hospital? Which software version? Which time period? Which economic conditions?

Generalisability asks what happens when the claim moves beyond the original setting. Sometimes the mechanism is stable. Sometimes the context changes everything.

A study of adults does not automatically answer a question about children. A trial in one health system may not transfer unchanged to another. A model trained on one distribution may fail after conditions shift. A teaching intervention that works with expert implementation may weaken when scaled.

The world is allowed to contain interactions.

27. Incentives Can Bend the Research System

Researchers do not operate outside institutions. Careers, grants, prestige, publication, deadlines and competition affect behaviour.

If surprising positive findings are easier to publish than careful null results, the visible literature can become distorted. If novelty is rewarded more strongly than verification, replication may be underproduced. If researchers face pressure to produce many papers quickly, methodological corners can become tempting.

Metascience studies these system-level failure modes. The important lesson is that research quality depends not only on individual integrity but also on the incentives and infrastructures surrounding researchers.

28. Open Science Tries to Make More of the Route Inspectable

Open data, open code, preregistration, registered reports, protocol sharing and open-access publishing are different interventions aimed at different parts of the research chain.

They should not be treated as magic labels. Sensitive data cannot always be made public. Open code can still contain errors. A preregistered study can still be badly designed.

The deeper objective is inspectability: make enough of the evidence-producing route visible that competent others can understand what happened, reproduce what is reproducible, challenge what is challengeable and learn from failure.

29. Qualitative Research Solves a Different Kind of Visibility Problem

Not every scientifically important question is best reduced immediately to a number.

If a new policy fails, researchers may need to understand how people interpreted it, where implementation broke, which meanings differed across groups, and what mechanisms were invisible to the original designers.

Interviews, observations, document analysis and other qualitative approaches can reveal structure that a pre-specified quantitative instrument might miss. Their rigor comes from transparent sampling, careful data collection, disciplined coding or interpretation, reflexivity, triangulation and an explicit account of how claims were derived.

The false choice is “numbers or stories”. Strong research asks which representation can best expose the phenomenon and which additional method could challenge the first representation.

30. Models and Simulations Are Research Laboratories for Assumptions

Sometimes researchers cannot directly observe the whole system or safely manipulate it. Models and simulations make assumptions executable.

They can explore climate futures, epidemic dynamics, traffic flows, molecular interactions, population change, engineering stress and many other complex systems.

But a model does not become reality because its output looks detailed. Its usefulness depends on assumptions, parameter estimates, calibration, validation, sensitivity and the purpose for which it is being used. How Models Work and How Simulation Works cover that boundary.

31. Worked Example: Does a Teaching Intervention Improve Learning?

Suppose a school introduces retrieval practice and wants to know whether it improves long-term learning.

A weak research plan might compare this year’s class with last year’s class, notice higher scores and announce success.

A stronger plan asks what else changed. Were the cohorts equivalent? Was the assessment identical? Did teachers gain experience? Was there additional tuition? Were students selected differently? Did attendance change? Did the intervention actually occur as intended?

Then it specifies the outcome. Immediate quiz performance is not the same as long-term retention. Familiar practice items are not the same as transfer. A student remembering after ten minutes is not the same as remembering next month.

Now the study might use random assignment where feasible, pre-specified outcomes, delayed tests, comparable instruction time, fidelity checks and an analysis plan that estimates both the magnitude and uncertainty of the effect.

Even after a positive result, the next questions remain: Does it work with other teachers? Other ages? Other subjects? Does the benefit persist? Are there costs or harms? Does it crowd out something else?

Research is not the act of making the first answer sound final. It is the construction of increasingly difficult opportunities for the answer to fail.

32. Worked Example: Why a City Feels Hotter

Suppose residents report that one neighbourhood feels much hotter than another.

The phenomenon can be studied at several levels.

• Air temperature sensors can measure physical conditions.
• Surface temperature imagery can reveal heat stored in materials.
• Tree-canopy mapping can reveal shading differences.
• Building geometry can alter wind and radiation.
• Surveys can capture perceived thermal comfort.
• Health records may reveal heat-related outcomes.

No single layer owns the whole explanation. The human experience of heat is related to, but not identical with, one thermometer reading. A strong programme of research connects measurements rather than forcing them into one proxy.

33. Worked Example: A Strange Biological Result

A laboratory observes that a compound appears to slow cell growth.

The first result is a clue, not a medicine.

Researchers may repeat the assay, test dose-response relationships, check contamination, use alternative measurement methods, compare different cell lines, investigate mechanism, test toxicity, move into animal models where ethically justified, and eventually—if the evidence remains promising—enter carefully regulated human studies.

At every stage, the claim changes resolution. “This signal changed in this assay” is a much smaller claim than “this treatment improves patient outcomes”. Research progress often consists of refusing to jump across those missing layers.

34. The Hostile Test: A Beautiful Paper With a Broken Chain

Imagine a paper with elegant figures, advanced statistics and a striking conclusion.

Now inspect the chain.

• The sample came from a narrow convenience group.
• The outcome changed after the researchers saw the data.
• Several analyses were tried but only one was reported.
• The key measure was never validated for this population.
• Thirty percent of participants disappeared from follow-up.
• The code cannot recreate the final figure.
• The conclusion generalises to people never represented in the study.

The paper may still look sophisticated.

But sophistication at the final layer cannot repair broken provenance earlier in the chain.

35. What Strong Research Reporting Makes Visible

• The exact research question.
• How the question relates to prior evidence.
• Which parts were exploratory and which confirmatory.
• The target population or system.
• The sampling or selection process.
• The operational definitions and measurement procedures.
• The study design and comparison logic.
• Ethical approvals and protections where relevant.
• Exclusions, missing data and deviations from plan.
• Analytic methods and assumptions.
• Effect sizes and uncertainty.
• Alternative explanations and limitations.
• Enough methods, data or code to support appropriate checking.
• The boundary beyond which the conclusion should not travel without new evidence.

This is why methods sections matter. They are not bureaucratic residue. They are the recoverable route between the claim and the world.

36. How to Read a Research Headline Without Being Fooled

When a headline says “Scientists find that X causes Y”, ask:

1. What study design was used?
2. Who or what was studied?
3. How was X measured?
4. How was Y measured?
5. Was X manipulated or merely observed?
6. What alternatives could explain the association?
7. How large was the effect?
8. How uncertain was it?
9. Was this an exploratory analysis?
10. Has the result been independently replicated?
11. Does the claim in the headline match the claim in the paper?
12. Does the result matter in the real world?

You do not need a doctorate to ask those questions. They are part of scientific literacy.

37. Primary School: Learning the Habit Before the Machinery

Young learners do not need the whole statistical apparatus at once. They need the intellectual habits that later research machinery protects.

• Notice carefully.
• Ask what changed.
• Compare fairly.
• Measure consistently.
• Record what actually happened.
• Do not change the story because the result is disappointing.
• Try again.

A child who learns to separate “I expected this” from “I observed this” has already crossed an important scientific threshold.

38. Secondary School: Learning to Separate Claim From Evidence

Secondary students can begin asking harder questions.

• What variable changed?
• What stayed controlled?
• What was the comparison?
• What evidence would weaken this explanation?
• Is the conclusion causal or correlational?
• Could measurement error create the pattern?
• Does the sample support the generalisation?

This is the stage where “scientific method” should stop being a memorised acronym and become a way of interrogating claims.

39. JC, University and Beyond: Research Becomes Argument Under Constraint

At higher levels, methods become specialised, but the governing problem remains recognisable.

A researcher is making an argument:

Given this question, these observations, these methods, these assumptions and this uncertainty, this is the conclusion the evidence currently supports.

The expertise lies partly in knowing how each word in that sentence can fail.

40. The Deep Structure: Research Is Organised Doubt

People sometimes describe science as a body of facts.

It is also an infrastructure for disciplined disagreement.

Different researchers can propose competing explanations. They can criticise one another’s methods. They can inspect data. They can collect new evidence. They can build better instruments. They can discover that a famous result was wrong. They can discover that an apparent failure to replicate reveals a previously hidden condition.

Scientific progress does not require scientists to become free of bias, ego, error or institutional pressure. It requires enough structure that those human limitations can be challenged by other people and, eventually, by reality itself.

41. Where Scientific Research Fits in the eduKateSG “How Works” Landscape

This article sits above several neighbouring mechanisms:

• How Science Works — the larger knowledge-building system.
• How Observation Works — how encounters with the world become records.
• How Measurement Works — how bounded attributes become comparable evidence.
• How Evidence Works — how observations bear on claims.
• How Models Work — how simplified representations support explanation and prediction.
• How Probability Works — how possible outcomes and information become calibrated belief.
• How Uncertainty Works — how limits remain visible.
• How Verification Works — how claims and requirements are independently checked.
• How Publishing Works — how research enters the durable public record.
• How Citation Works — how evidence lineage can be followed.

Scientific research connects these mechanisms into one practical route from uncertainty to a claim that other people can challenge.

42. What This Article Does Not Claim

• Science does not follow one universal sequence of steps.
• An experiment is not required for every scientific question.
• A peer-reviewed paper is not automatically correct.
• A statistically significant result is not automatically important.
• A non-significant result does not automatically prove no effect.
• A large sample does not automatically remove bias.
• A randomised trial does not automatically generalise everywhere.
• Replication failure does not automatically identify which study is wrong.
• Open data does not override legitimate privacy, consent or security constraints.
• Research cannot eliminate uncertainty. It can make uncertainty more explicit, bounded and contestable.

43. A Compact Research Audit

1. What exactly is the question?
2. Why does it matter?
3. What is already known?
4. What are the serious alternative explanations?
5. What population or system is the claim about?
6. How were cases sampled or selected?
7. What was measured and how?
8. What comparison makes the inference possible?
9. What biases could enter?
10. What assumptions does the analysis require?
11. How large is the estimated effect or relationship?
12. How uncertain is it?
13. Were the analyses pre-specified or exploratory?
14. What evidence would change the conclusion?
15. Can the computational route be reproduced?
16. Has the result survived new data?
17. Does it fit the wider body of evidence?
18. Where should the conclusion stop?

44. Frequently Asked Questions

What is scientific research in one sentence?

Scientific research is disciplined inquiry that uses transparent methods, evidence and criticism to produce claims that remain open to independent checking and correction.

Does every study need a hypothesis?

No. Hypothesis testing is central to many studies, but descriptive, exploratory, qualitative, methodological and theory-building research can be scientifically valuable without beginning from one formal hypothesis.

Is a randomised controlled trial the best research design?

It is extremely powerful for some causal questions, but “best” depends on the inference. Many questions cannot be randomised ethically or practically, and randomisation does not automatically solve measurement, attrition, implementation or generalisability problems.

What is the difference between reproducibility and replication?

Following the National Academies terminology, reproducibility concerns obtaining consistent results from the same data and computational route, while replication involves new data addressing the same scientific question.

Does peer review prove a paper is true?

No. Peer review can detect problems and strengthen manuscripts, but it is one filter in a longer process of criticism, reproduction, replication and synthesis.

Why can scientists disagree?

They may have different data, methods, assumptions, models or interpretations. Productive disagreement becomes scientific when those differences are made explicit enough that evidence can adjudicate among them.

Why does scientific knowledge change?

Because new observations, better instruments, stronger designs, replication attempts and improved theories can change the balance of evidence. Correctability is a feature of the system, not an admission that knowledge is arbitrary.

45. Authoritative Research Corridor

• National Academies / NCBI Bookshelf — Scientific Methods and Knowledge
• National Academies / NCBI Bookshelf — Understanding Reproducibility and Replicability
• National Academies / NCBI Bookshelf — Replicability
• Nature Biomedical Engineering — Methods, Techniques, Assays and Protocols
• Nature Human Behaviour — A Manifesto for Reproducible Science
• International consensus on core reproducibility items in research

Final Thought: Research Is a Way of Making Error Findable

The most impressive research result is not the one that sounds most certain.

It is the one whose route from question to evidence is clear enough that another competent person can inspect it, challenge it, repeat the relevant parts, collect new evidence and—if necessary—show where it fails.

That is why science can progress even though scientists are human.

The system does not require every researcher to be right on the first attempt.

It requires claims to remain vulnerable to better observation.

A question enters.

A method is built.

Evidence pushes back.

Other people check.

The answer changes.

And the world gets another chance to say no.