How Scientific Hypotheses Work | From a Testable Idea to a Claim That Can Survive Failure

A scientific hypothesis is a claim that observations or experiments can test. This guide explains how to turn a question into a precise idea, derive predictions, consider alternatives and decide what evidence would support the claim or lead you to revise it.

A scientific hypothesis is not merely a guess, and it is not automatically a theory in miniature. It is a candidate claim or explanatory proposition framed strongly enough that observations or experiments could count for it, against it, or force its revision. The central discipline is not confidence. It is testability: the hypothesis must make contact with possible evidence that could have turned out differently.

Work through the examples from a question to a candidate explanation, a prediction and a test. Related guides explore scientific models, evidence, experiments, explanations and prediction. Together, they help you separate an idea that sounds plausible from one that can be examined critically.

---

Choose your route

• Basic route: what a hypothesis is, what makes it testable, and how it differs from a guess.
• School Science route: question → hypothesis → prediction → experiment → evidence.
• Research route: alternative hypotheses, auxiliary assumptions, operationalisation, confounders and replication.
• Statistics route: null hypotheses, alternatives, p-values, effect sizes, Type I/II errors and why rejection is not proof.
• Theory route: how hypotheses arise from theories, observations, models and anomalies.
• Advanced route: underdetermination, falsification, severe tests, HARKing, preregistration and model comparison.

1. A hypothesis is a candidate claim exposed to evidence

Science begins with questions, but questions alone do not tell us what to test.

Suppose a pupil notices that one plant grows faster than another.

The question might be:

“Why did Plant A grow faster?”

Possible hypotheses include:

• Plant A received more light.
• Plant A received more water.
• Plant A began larger.
• Plant A belongs to a different variety.
• Plant A’s soil differed.

Each hypothesis points toward different evidence.

The hypothesis is useful because it narrows a broad question into something that can generate a test.

2. A hypothesis is not just any sentence

“Plants are interesting” is not a scientific hypothesis.

“This plant likes music” is too vague unless “likes” is translated into something observable or measurable.

“Under otherwise comparable conditions, plants exposed to treatment X will show greater measured growth over four weeks than untreated plants” is much closer to a scientific hypothesis because:

• the comparison is defined;
• the outcome is measurable;
• the condition is stated;
• different results are possible.

A hypothesis becomes scientifically useful when its language creates a bridge to evidence.

3. Testability is the centre

A claim is scientifically testable when we can specify observations that would bear on it.

Ask:

• What would we expect to observe if this idea is correct?
• What would we expect if it is wrong?
• Can those observations actually be obtained?
• Could the result have turned out differently?

The University of California, Berkeley’s Understanding Science project emphasises that testable ideas sit at the centre of science and that real scientific inquiry is more flexible than the oversimplified classroom “five-step method”.

4. Falsifiability: the possibility of being wrong

A hypothesis gains scientific force when it excludes some possible outcomes.

If every conceivable observation can be reinterpreted as support, the hypothesis is insulated from evidence.

Consider:

“An invisible force affects the experiment, but whenever you try to measure it, it changes so the result always looks exactly as if no force exists.”

If no observation could ever count against the claim, the claim is difficult to test scientifically.

Falsifiability does not mean scientists must expect a hypothesis to be false.

It means the hypothesis leaves open the possibility that experience could show it is wrong.

5. Falsification is not always simple

If a prediction fails, the hypothesis may be wrong.

But so might:

• the instrument;
• the measurement procedure;
• an auxiliary assumption;
• the implementation of the experiment;
• the sample;
• a background model used to derive the prediction.

Scientific diagnosis therefore asks which part of the reasoning chain failed.

A failed prediction is evidence against the full package used to generate it, not always a magical one-step proof that one sentence is false.

6. Hypothesis versus prediction

A hypothesis proposes a relationship, mechanism or state of affairs.

A prediction is an expected observation that follows if the hypothesis and relevant assumptions are correct.

Example:

Hypothesis: increased light availability increases measured growth over this interval, within the tested range and under comparable other conditions.

Prediction: plants assigned to the higher-light condition will show greater mean measured growth than plants in the lower-light condition.

Prediction converts the hypothesis into something the experiment can confront.

7. One hypothesis can generate many predictions

A strong hypothesis is often fertile.

If a mechanism is real, it may imply:

• a directional change;
• a dose-response relationship;
• a time-course pattern;
• a boundary condition;
• a different response in another population;
• a change under intervention.

Multiple predictions can test different parts of the same underlying idea.

8. One prediction can be compatible with several hypotheses

Suppose a treatment group improves.

Possible explanations might include:

• the treatment had the intended causal effect;
• groups differed at baseline;
• measurement changed;
• participants changed behaviour because they knew they were treated;
• another factor differed between groups.

A prediction matching reality does not automatically identify which hypothesis is correct.

Good experimental design attempts to discriminate among alternatives.

9. Alternative hypotheses are not afterthoughts

A weak investigation asks:

“How can I prove my idea?”

A stronger investigation asks:

“What other hypotheses could explain the same observation, and what evidence would distinguish them?”

Alternative hypotheses make tests sharper.

They reduce confirmation bias because the scientist must consider what else could be true.

10. Generate alternatives before testing

Before collecting data, list plausible explanations.

For a plant-growth difference:

• light;
• water;
• starting size;
• soil;
• genetics;
• measurement error.

Then ask which design decisions make one explanation more distinguishable from the others.

This is the intellectual foundation of controls.

11. The hypothesis should be specific enough to risk failure

Weak:

“Temperature affects the reaction.”

Stronger:

“Within the tested range, increasing temperature will increase the measured reaction rate under otherwise comparable conditions.”

The stronger statement:

• identifies direction;
• defines a domain;
• implies measurable outcome;
• can fail.

12. But do not make the hypothesis artificially specific

False precision is not rigour.

If prior evidence does not justify predicting an exact 17.4% increase, do not invent that number merely to look scientific.

Specificity should come from theory, prior data or a model.

13. Directional versus non-directional hypotheses

A directional hypothesis predicts the direction of an effect.

Example:

“Treatment A will produce a larger mean improvement than Treatment B.”

A non-directional hypothesis predicts a difference without specifying which direction.

Example:

“Treatment A and Treatment B will produce different mean outcomes.”

Directional claims are narrower. They should be justified before seeing the data rather than chosen afterward.

14. A hypothesis needs a domain

Statements that sound universal often depend on conditions.

Instead of:

“Higher temperature always increases reaction rate.”

Ask:

• within what temperature range?
• for which reaction?
• under what pressure?
• with which concentrations?

Scientific hypotheses become stronger when their scope is explicit.

15. Operationalisation: turn ideas into things we can observe

Many scientific concepts are not measured directly.

Researchers must decide how to represent them operationally.

Examples:

• “growth” might mean change in height, mass or leaf area;
• “stress” might be represented by questionnaires, physiology or behaviour;
• “fitness” can mean different biological quantities;
• “learning” can be measured by retention, transfer or test performance.

The operational definition is not identical to the underlying concept.

It is the measurement bridge.

16. Bad operationalisation can sink a good hypothesis

A hypothesis may be sensible while the chosen measure is poor.

If the measurement does not capture the construct well, the test can produce misleading results.

Always ask:

• Does this measurement represent the idea?
• Is it reliable?
• Is it sensitive enough?
• Does it create bias?

17. Hypotheses can arise from theory

A mature theory can generate hypotheses about new conditions.

The chain is:

theory → hypothesis → prediction → test → evidence → theory update

This is often called a hypothetico-deductive structure.

The hypothesis acts as a testable consequence or narrower proposition derived from the larger framework.

18. Hypotheses can arise from observation

Science is not always theory-first.

An unexpected pattern can generate a hypothesis.

Example:

A researcher notices that one population behaves differently under the same apparent conditions.

Possible hypotheses are generated to explain the anomaly.

Then new evidence is sought.

19. Hypotheses can arise from failed predictions

A model predicts X.

The world produces Y.

The discrepancy can generate a new hypothesis:

“Perhaps the model omitted Process Z.”

Science therefore cycles:

prediction failure → hypothesis generation → new test → model revision

20. Hypotheses can arise from analogy

Scientists often ask:

“If this mechanism operates in System A, might a related mechanism operate in System B?”

Analogy can generate useful hypotheses.

But analogy is not evidence by itself.

The new system must be tested on its own terms.

21. Hypotheses can arise from computation

Large datasets can reveal unexpected clusters or associations.

These patterns can generate hypotheses.

But if the same data are used both to discover and confirm the idea, the evidence is less independent.

A healthy sequence is:

discover pattern → form hypothesis → test on new data

22. Exploration and confirmation are both science

Exploratory work is not inferior.

It is how many discoveries begin.

The problem arises when exploratory findings are presented as if they were predicted in advance.

Science becomes clearer when we label the stage honestly.

23. HARKing

HARKing means Hypothesizing After the Results are Known.

The problematic version occurs when researchers inspect the result, create a hypothesis to fit it, and present that hypothesis as if it had been specified beforehand.

This can inflate apparent confirmation because the hypothesis was designed with access to the answer.

A better response is:

“This result suggests a new hypothesis, which should now be tested independently.”

24. Preregistration

Preregistration records key elements of a study before outcomes are known.

It can include:

• research question;
• hypotheses;
• outcomes;
• sample plan;
• analysis plan;
• exclusion rules.

Preregistration does not eliminate exploration.

It helps distinguish confirmatory analysis from exploratory analysis.

25. Registered reports

In a registered report, study rationale and methods are reviewed before results are known.

This reduces publication incentives that favour surprising positive results after the fact.

The approach treats good questions and methods as scientifically valuable regardless of whether the hypothesis is supported.

---

26. Null hypotheses

In statistical testing, a null hypothesis specifies a statistical state against which data are compared.

A common example is:

“There is no difference in the population means.”

But null hypotheses need not always mean literal “no effect”.

They can specify a particular value, equivalence boundary or other reference state.

27. Statistical hypothesis testing is not the whole meaning of hypothesis

A scientific hypothesis can be a mechanistic or causal idea.

A statistical null hypothesis is a formal object in statistical inference.

Do not collapse the two.

The scientific question should determine the statistical formulation, not the reverse.

28. Null and alternative hypotheses

In many standard tests:

H₀: specifies the null condition.

H₁ or Hₐ: specifies an alternative.

The data are used to evaluate whether the observed statistic is sufficiently inconsistent with H₀ under the test’s assumptions.

29. Rejecting the null does not prove the alternative

A small p-value can justify rejecting a particular null model under stated assumptions.

It does not prove:

• the research hypothesis is certainly true;
• the effect is large;
• the mechanism is correct;
• the study is unbiased;
• the result will replicate.

Statistical testing is one layer of evidence, not a machine for converting uncertainty into certainty.

30. Failing to reject is not proof of no effect

A non-significant result can occur because:

• the null is approximately true;
• the sample is small;
• measurement is noisy;
• the effect is smaller than the design can detect;
• assumptions are violated.

“No statistically significant evidence” is not automatically “evidence of no difference”.

31. Power

Statistical power is the probability that a test will detect a specified effect when that effect is present, under the model assumptions.

Low power makes negative results difficult to interpret.

Power depends on:

• sample size;
• effect size;
• measurement variability;
• test design;
• significance threshold.

32. Type I and Type II errors

Type I error: rejecting the null when it is true under the statistical model.

Type II error: failing to reject the null when the specified alternative is true.

Reducing one kind of error can increase the other unless sample size or information improves.

33. Statistical significance versus scientific significance

A tiny effect can be statistically significant in a huge sample.

A scientifically important effect can fail to reach conventional significance in a small noisy sample.

Always ask:

• How large is the estimated effect?
• How uncertain is that estimate?
• Does the magnitude matter scientifically or practically?

34. Effect size

Effect size describes magnitude.

Examples include:

• mean difference;
• risk ratio;
• correlation;
• standardised effect;
• odds ratio.

A hypothesis test without effect magnitude can leave readers knowing that a difference exists statistically but not whether it matters.

35. Confidence intervals

Confidence intervals provide a range of parameter values compatible with the data and model under a defined repeated-sampling procedure.

They can reveal:

• estimate precision;
• plausible effect magnitudes;
• whether small, null or large effects remain compatible.

They should be interpreted carefully rather than as a simple probability that the true value lies inside one realised interval.

---

36. Hypotheses should compete

Strong science often compares several hypotheses rather than testing one against a vague “everything else”.

Suppose a population falls.

Hypotheses:

• food became scarce;
• predation increased;
• disease increased;
• migration increased;
• measurement changed.

The next investigation should seek evidence that distinguishes among them.

37. Discriminating tests

A discriminating test is valuable because competing hypotheses predict different outcomes.

If Hypothesis A and Hypothesis B both predict the same observation, that observation does little to choose between them.

Ask:

“Where do the hypotheses disagree?”

Then design the test there.

38. Crucial experiments are rare

Textbooks sometimes describe one experiment that definitively chooses among theories.

Real scientific systems are often more complicated.

Experiments can depend on auxiliary assumptions, measurement models and background knowledge.

Evidence usually accumulates across multiple tests.

39. Converging evidence

A hypothesis gains strength when different methods, measurements and populations point in the same direction.

Convergence is especially persuasive when:

• methods have different biases;
• data sources are independent;
• predictions were specified in advance;
• replications succeed.

40. Independent lines of evidence reduce shared-error risk

Ten studies using the same instrument and same flawed assumption are not as independent as ten studies using different methods.

Evidence quality depends on diversity of failure modes, not only count.

41. Auxiliary hypotheses

Testing one central hypothesis often requires additional assumptions.

Example:

• the instrument measures what we think it measures;
• the treatment was delivered correctly;
• the sample is comparable;
• the statistical model is adequate.

These are auxiliary hypotheses.

A failed prediction may challenge the main hypothesis, an auxiliary hypothesis, or both.

42. The Duhem–Quine problem in practical language

Scientific tests often evaluate bundles of assumptions rather than one isolated statement.

This does not make testing impossible.

It means diagnosis requires follow-up tests.

When prediction fails, ask which assumption can be varied independently.

43. Hypothesis rescue can be legitimate

Suppose a failure reveals that a temperature sensor was miscalibrated.

Correcting the sensor and retesting is legitimate.

Science should repair known flaws.

44. Ad hoc rescue is different

If every failure is explained away with a new exception that makes no independent prediction, the hypothesis becomes protected rather than tested.

A useful modification should:

• have independent justification;
• improve future predictions;
• apply beyond the one failure;
• face new tests.

45. Hypotheses can be nested

A broad hypothesis can contain narrower versions.

Example:

“Temperature affects growth” can be refined into several hypotheses about range, threshold, direction or mechanism.

Nested hypotheses help scientists move from broad observation to precise testing.

46. Boundary hypotheses

Sometimes the key question is:

“Where does the rule stop working?”

Testing boundary conditions can be more informative than repeating another central example.

Boundaries reveal the domain of the hypothesis.

47. Mechanistic hypotheses

A mechanistic hypothesis proposes how the effect occurs.

It can generate deeper predictions than a simple association.

If A causes B through Process M, then manipulating or measuring M can produce additional tests.

48. Descriptive hypotheses

Not every hypothesis needs a detailed mechanism.

A descriptive hypothesis may propose:

• a pattern;
• a distribution;
• a relationship;
• a difference between groups.

Mechanistic understanding may come later.

49. Causal hypotheses

A causal hypothesis proposes that changing one factor would change another, under specified conditions.

Causal claims require stronger design than simple prediction.

Randomisation, natural experiments, causal models and careful control of confounding can help.

50. Correlational hypotheses

A correlational hypothesis proposes association without necessarily claiming cause.

It can be scientifically valuable.

But language should match the design.

Do not upgrade “associated with” into “causes” without causal evidence.

---

51. Hypotheses through the school years

The word “hypothesis” can be introduced too mechanically. Students are sometimes taught to write “If…, then…, because…” and may conclude that the grammar is the science. The real developmental job is deeper: learn to propose an idea that explains or predicts something, derive what should be observed if it is true, and then allow evidence to change the idea.

Primary 3: a prediction with a reason

At P3, the full research-language distinction between hypothesis and prediction may not yet be necessary. The pupil can begin with:

• What do you think will happen?
• Why do you think so?
• What did we observe?
• Did the observation support your idea?

This is the seed of hypothesis-driven thinking.

Primary 4: changed condition → expected effect

P4 can make the relationship clearer.

Example:

Question: What happens to evaporation if exposed surface increases while other relevant conditions are comparable?

Candidate idea: greater exposed surface increases the rate of evaporation under these conditions.

Prediction: the setup with greater exposed surface will show greater water loss over the same time interval.

The learner begins to distinguish the idea from the observation that tests it.

Primary 5: competing explanations

P5 should begin asking:

“What else could explain the result?”

For a plant-growth difference:

• light;
• water;
• starting size;
• soil;
• measurement.

A good experiment tries to reduce the number of plausible alternatives.

Primary 6: testable claims and evidence boundaries

P6 can be expected to:

• state a prediction before seeing the result;
• identify the factor being tested;
• recognise alternative explanations;
• explain why controls matter;
• revise conclusions when evidence disagrees;
• avoid saying a hypothesis is “proved”.

This creates a stronger runway into Secondary Science.

Secondary Science: explicit hypothesis–prediction distinction

Secondary pupils can work explicitly with:

hypothesis → expected relationship → prediction → experiment → result → evaluation

They can also learn that one failed result may challenge the hypothesis, the method, a measurement or an auxiliary assumption.

JC and advanced study: statistical and causal hypotheses

At higher levels, hypotheses become formal objects in statistical testing, causal inference and theory evaluation.

Students should learn to keep three layers separate:

1. the scientific question;
2. the substantive research hypothesis;
3. the statistical hypothesis used in analysis.

Confusing these layers creates many common research errors.

---

52. The “If–Then–Because” sentence: useful scaffold, dangerous prison

The familiar classroom form can help younger students:

If X changes, then Y will change, because Z.

But the scaffold has limits.

It can create several misconceptions:

• every hypothesis must be causal;
• every hypothesis must predict direction;
• the “because” clause is automatically a valid mechanism;
• writing the sentence correctly makes it scientifically good.

The content matters more than the syntax.

53. Use scaffolds, then fade them

Early:

“If light increases, then measured growth will increase, because…”

Later:

“State a testable hypothesis about the relationship between light and growth over the tested range.”

Later still:

“Propose two competing explanations and a test that discriminates them.”

The goal is independent reasoning, not permanent template dependence.

54. A good hypothesis creates risky predictions

Suppose two hypotheses explain existing data equally well.

Hypothesis A predicts:

“Under Condition C, outcome increases.”

Hypothesis B predicts:

“Under Condition C, outcome does not change.”

Condition C is scientifically valuable because the hypotheses disagree there.

A risky prediction narrows the expected future enough that one model can lose.

55. A hypothesis can survive without being proven

If repeated tests support the predictions, confidence can increase.

But future evidence can still reveal:

• a boundary;
• a confounder;
• a missing mechanism;
• a restricted domain;
• a better competing explanation.

Scientific support is often provisional.

56. “Supported” is often better than “proved”

Strong scientific language includes:

• supports;
• is consistent with;
• provides evidence for;
• fails to support;
• is inconsistent with;
• requires revision.

These phrases preserve the relationship between evidence and uncertainty.

---

57. Hypothesis Casebook 1 — Plants and music

A student asks whether music helps plants grow.

Weak hypothesis:

“Plants like music.”

Problems:

• “like” is not operationalised;
• music can differ in volume, duration and type;
• growth needs a measurement;
• other conditions can differ.

Stronger hypothesis:

“Under otherwise comparable conditions, plants exposed to Sound Treatment X for one hour daily will show a different mean increase in height over four weeks than plants not exposed to that treatment.”

This is testable, but it still does not specify mechanism.

A mechanistic follow-up could ask whether vibration, stress or another process is involved.

58. Hypothesis Casebook 2 — Fertiliser and growth

Observation:

Plants receiving fertiliser are taller.

Possible hypotheses:

• fertiliser caused increased growth;
• fertilised plants also received more water;
• starting sizes differed;
• the fertilised group had better light;
• measurement differed.

A randomised controlled design can reduce some alternatives.

The teaching point is that an observed difference is not identical to one causal explanation.

59. Hypothesis Casebook 3 — Salt and boiling point

Question:

Does dissolved salt change the boiling temperature of water?

A testable hypothesis can specify:

• salt concentration;
• measured boiling temperature;
• pressure or altitude conditions;
• comparison with unsalted water.

This case illustrates domain control. Boiling temperature also depends on pressure, so environmental assumptions matter.

60. Hypothesis Casebook 4 — Exercise and heart rate

Hypothesis:

Increasing exercise intensity increases heart rate within the tested range.

Alternative explanations for differences among participants include:

• fitness;
• age;
• medication;
• measurement timing;
• hydration;
• stress.

A repeated-measures design can control some between-person differences.

61. Hypothesis Casebook 5 — Studying and test scores

Hypothesis:

More retrieval practice improves delayed test performance compared with rereading, under defined conditions.

Potential confounders:

• total study time;
• prior knowledge;
• question difficulty;
• motivation;
• feedback quality.

The hypothesis becomes meaningful only after “study method” and “learning” are operationalised.

62. Hypothesis Casebook 6 — Phone use and sleep

Observation:

High evening phone use correlates with shorter sleep.

Possible hypotheses include:

• phone use delays bedtime;
• light exposure affects sleep timing;
• stress causes both phone use and poor sleep;
• people unable to sleep use phones more.

This case is useful because causal direction is ambiguous.

A correlational study can identify association, but stronger causal claims require better design.

63. Hypothesis Casebook 7 — Pollution and asthma

A population study finds an association between pollution exposure and asthma symptoms.

Possible confounders include:

• housing;
• income;
• smoking exposure;
• occupation;
• access to healthcare.

The hypothesis may be scientifically important even when randomised exposure is unethical.

Science must sometimes rely on natural experiments, longitudinal designs and causal modelling.

64. Hypothesis Casebook 8 — Drug efficacy

Scientific hypothesis:

The drug improves a defined clinical outcome compared with control.

Statistical null:

The population effect equals the specified null value.

These are related but not identical.

The trial must also evaluate:

• effect size;
• confidence interval;
• adverse events;
• population applicability;
• replication.

65. Hypothesis Casebook 9 — Astronomy

A new astronomical signal appears periodically.

Alternative hypotheses might include:

• an orbiting object;
• instrumental artifact;
• stellar variability;
• data-processing artifact.

Different hypotheses predict different timing, wavelengths or independent observations.

Follow-up measurement should target the points where the alternatives diverge.

66. Hypothesis Casebook 10 — Fossil interpretation

A new fossil has a mixture of traits.

Hypotheses can concern:

• evolutionary relationship;
• age;
• environment;
• function;
• developmental stage.

Testing may use comparative anatomy, dating, geology and phylogenetic analysis.

Science does not require every hypothesis to be tested by a laboratory experiment.

67. Hypothesis Casebook 11 — Climate attribution

Question:

What explains an observed long-term temperature trend?

Candidate hypotheses involve:

• greenhouse-gas forcing;
• solar variation;
• volcanic forcing;
• internal variability;
• measurement artifacts.

Scientists compare model fingerprints, observations and multiple lines of evidence.

The hypothesis-testing architecture can operate at planetary scale.

68. Hypothesis Casebook 12 — Machine failure

A machine begins failing more often.

Hypotheses:

• component wear;
• temperature increase;
• new supplier material;
• sensor drift;
• software update;
• maintenance change.

Logs and controlled replacement tests can discriminate causes.

The same scientific reasoning appears in engineering troubleshooting.

69. Hypothesis Casebook 13 — A model suddenly performs worse

Possible hypotheses:

• distribution shift;
• sensor change;
• data pipeline bug;
• label definition changed;
• model degradation;
• user behaviour changed.

Testing should vary one layer at a time where possible.

Hypothesis generation is central to debugging predictive systems.

70. Hypothesis Casebook 14 — Unexpected school result

A student who usually performs well suddenly scores poorly.

Possible hypotheses:

• topic gap;
• time pressure;
• question-reading failure;
• fatigue;
• careless checking;
• an unusually difficult paper.

A diagnostic tutor should not jump immediately to “student did not study”.

Multiple hypotheses should be compared against evidence.

71. Hypothesis Casebook 15 — Why did the experiment fail to replicate?

Possible hypotheses:

• original effect was a false positive;
• replication sample differs;
• protocol differs;
• measurement differs;
• effect depends on context;
• original estimate was exaggerated.

Replication failure generates new hypotheses rather than one automatic conclusion.

---

72. Hypothesis quality checklist

A strong hypothesis should answer as many of these as appropriate:

• What relationship or mechanism is proposed?
• Which variables or entities are involved?
• What population or system does it apply to?
• What domain or conditions matter?
• What observation would support it?
• What observation would count against it?
• What competing hypotheses exist?
• What prediction follows?
• Can the key terms be measured or operationalised?
• Is the claim more specific than the evidence justifies?

73. A hypothesis should guide evidence collection

If writing the hypothesis does not change what evidence you would collect, it may be too vague.

Useful hypotheses influence:

• what is measured;
• which groups are compared;
• what controls are needed;
• which time window matters;
• what prediction is made;
• which analysis is appropriate.

74. Do not confuse a research question with a hypothesis

Question:

“How does temperature affect enzyme activity?”

Hypothesis:

“Within Range R, activity will increase with temperature until Condition C, after which activity will decrease because…”

The question opens possibilities.

The hypothesis commits to one of them.

75. Do not force hypotheses when the study is exploratory

Some research legitimately asks:

• What species live here?
• What patterns exist in this dataset?
• What is the structure of this molecule?
• What anomalies appear?

Not every useful scientific investigation begins with a formal hypothesis.

Exploration can generate the hypotheses that later studies test.

---

76. Replication: a hypothesis must survive outside its first success

One successful experiment is rarely the end of the story.

A result can be influenced by:

• chance;
• measurement;
• sample peculiarities;
• analysis choices;
• context;
• unnoticed bias.

Replication asks whether the result or relationship appears again under a new test.

77. Direct replication

A direct replication attempts to reproduce the original study as closely as practical.

The question is:

“If we repeat the same essential procedure with a new sample, do we observe a compatible result?”

Direct replication tests repeatability of the claimed effect under similar conditions.

78. Conceptual replication

A conceptual replication tests the same underlying hypothesis using a different method, measurement or context.

This is scientifically powerful because it changes the failure modes.

If the same relationship appears when measured differently, confidence can increase that the effect is not merely an artifact of one procedure.

79. Replication and domain

A failed replication does not always mean the original hypothesis is universally false.

It can reveal that the effect depends on:

• population;
• temperature;
• culture;
• dosage;
• time window;
• measurement;
• another boundary condition.

Replication can refine the domain of a hypothesis.

80. Reproducibility and replicability are different jobs

In modern research discussions, reproducibility often refers to obtaining consistent computational or analytic results from the same data, code and procedures.

Replicability often refers to obtaining consistent scientific results in a new study that collects new data.

The exact terminology varies across fields, so authors should define how they use the terms.

81. Computational reproducibility

A result becomes difficult to audit if no one can reconstruct:

• the data used;
• the preprocessing steps;
• the code;
• the model version;
• the random seed;
• the analysis decisions.

Reproducible workflows help separate scientific disagreement from hidden procedural differences.

82. Open materials

Where ethical, legal and practical constraints allow, research can be strengthened by sharing:

• protocols;
• analysis code;
• materials;
• synthetic or de-identified data;
• model specifications;
• preprocessing details.

Transparency makes hypotheses easier to test independently.

83. Replication is not punishment

A replication is not an accusation against the original researchers.

It is part of normal science.

Important claims deserve repeated testing precisely because people care whether they are reliable.

84. The first published effect can be too large

Early studies are sometimes selected for unusually large or exciting results.

Later studies can estimate smaller effects.

This can happen because:

• initial samples were small;
• publication favoured significant findings;
• measurement improves;
• the true effect is heterogeneous;
• chance inflated the first estimate.

A shrinking effect is not necessarily evidence of misconduct.

It may be the normal correction of uncertainty.

85. Publication bias

Studies with striking positive results can be easier to publish than studies with null or ambiguous results.

If the literature contains mainly successful tests of a hypothesis, readers may overestimate support.

Publication bias affects the evidence base even when every published study was conducted honestly.

86. The file-drawer problem

Imagine 20 laboratories test a weak hypothesis.

By chance, one produces a statistically significant result.

If only that one study is published, the literature can look much stronger than reality.

Registration of studies and publication of null results can reduce this distortion.

87. Selective outcome reporting

A study may measure many outcomes and report only the ones that support the hypothesis.

This is a form of selection.

Preregistration helps make the planned primary outcomes visible.

Exploratory outcomes can still be reported, but should be labelled accordingly.

88. Multiple comparisons

If researchers test many hypotheses, some will appear significant by chance.

Example:

If 100 independent null hypotheses are tested at a 5% significance level, around five false positives are expected on average under idealised assumptions.

Multiple-testing procedures can control error rates, but the best solution also includes clear hypothesis planning and replication.

89. Researcher degrees of freedom

Researchers often make many reasonable choices:

• which participants to exclude;
• which outcome transformation to use;
• which covariates to include;
• which subgroup to analyse;
• where to stop collecting data.

If these decisions are influenced by the results, false-positive risk can rise.

90. P-hacking

P-hacking refers to using analysis flexibility to search for statistically significant results and then presenting the selected result as if it came from one planned test.

Examples can include:

• trying many outcomes;
• trying many exclusion rules;
• stopping data collection after significance appears;
• testing many subgroups without correction.

Hypothesis transparency helps reduce this problem.

91. Garden of forking paths

Even without deliberate p-hacking, analysis can contain many branches.

Each decision depends on what the data look like.

The final analysis may therefore have been selected through a hidden path of possibilities.

Preregistration, multiverse analysis and robustness checks can reveal this flexibility.

92. Multiverse analysis

A multiverse analysis examines many reasonable analytic choices rather than reporting only one.

If the conclusion remains similar across choices, robustness increases.

If the result changes dramatically, the hypothesis may depend on a narrow analytic path.

93. Sensitivity analysis

Sensitivity analysis asks how the conclusion changes when assumptions change.

Examples:

• different inclusion criteria;
• different model specifications;
• different measurement thresholds;
• different missing-data methods.

A robust hypothesis should not collapse under tiny reasonable changes unless the system itself is truly near a threshold.

---

94. Bayesian hypothesis updating

Bayesian reasoning treats hypotheses as having prior probabilities or plausibilities that update after evidence.

The basic structure is:

prior → likelihood of evidence under each hypothesis → posterior

This makes alternative hypotheses explicit.

95. Evidence is comparative

Evidence supports Hypothesis A over Hypothesis B when the observation is more expected under A than under B.

This is deeper than asking whether the evidence is “compatible” with A.

Many observations are compatible with many hypotheses.

The key question is:

Which hypothesis predicted the evidence better?

96. Likelihood ratios

A likelihood ratio compares how probable the observed evidence is under two hypotheses.

If evidence E is ten times more likely under A than B, E shifts support toward A.

This does not by itself give the posterior probability unless prior odds are also considered.

97. Priors are not automatically subjective guesses

Priors can reflect:

• previous studies;
• physical constraints;
• base rates;
• expert knowledge;
• weakly informative assumptions.

They should be stated and sensitivity-tested when conclusions depend strongly on them.

98. Bayes factors

A Bayes factor compares how well two statistical hypotheses predict the observed data, integrated over their parameter distributions.

It can provide evidence for one hypothesis over another rather than only testing a null for rejection.

Interpretation still depends on model specification and priors.

99. Posterior probability is model-conditional

A posterior probability answers a question within the considered model set.

If all candidate hypotheses omit the true explanation, posterior certainty can still be misleading.

Always ask whether important alternatives are missing.

100. Bayesian and frequentist methods answer different questions

Neither framework should be reduced to caricature.

Frequentist methods often ask about the behaviour of procedures under repeated sampling.

Bayesian methods update probabilities over hypotheses or parameters under a specified prior-model structure.

The scientific question should guide the method.

---

101. Equivalence testing

Sometimes the scientific hypothesis is not “there is a difference”.

It is:

“Any difference is smaller than a scientifically meaningful threshold.”

Equivalence testing can address this question directly.

This is stronger than merely failing to reject a zero-effect null.

102. Non-inferiority hypotheses

In some trials, the goal is to show a new intervention is not unacceptably worse than a standard by more than a pre-specified margin.

The scientific logic depends heavily on the chosen margin.

The margin should be justified before results are known.

103. Minimal effect tests

Researchers may want to test whether an effect exceeds a meaningful minimum size.

This shifts attention from “different from zero” to “large enough to matter”.

Scientific hypotheses should reflect meaningful questions rather than default statistical rituals.

104. Prediction intervals versus confidence intervals

A confidence interval usually concerns an unknown parameter.

A prediction interval concerns a future observation.

They answer different questions.

Hypotheses about mechanisms and forecasts should use the interval suited to the target.

105. Sequential testing

Some studies examine data as they accumulate.

If ordinary fixed-sample tests are repeatedly applied without adjustment, error rates can inflate.

Sequential designs use stopping rules designed for interim looks.

Again, planning before outcomes matters.

106. Adaptive experiments

Experiments can adapt:

• sample allocation;
• dose;
• arms;
• stopping.

Adaptation does not make a study unscientific.

But the statistical and decision rules should account for the adaptation.

107. Null results are scientifically useful

A well-designed study that fails to support a hypothesis can:

• rule out large effects;
• identify boundary conditions;
• challenge a theory;
• refine measurement;
• redirect research.

Science should not reward only positive results.

108. Negative evidence is strongest when the test was capable

If an experiment was too weak to detect the predicted effect, a null result says little.

Negative evidence becomes stronger when:

• measurement was sensitive;
• sample size was adequate;
• the predicted effect should have been visible;
• controls worked;
• the method was validated.

109. Severe testing intuition

A test is severe for a hypothesis when the hypothesis would likely have failed if it were wrong in the relevant way.

Passing an easy test provides less evidence than passing a demanding one.

Scientific confidence should reflect test severity, not merely the number of tests.

110. Confirmation from weak tests can be cheap

If a hypothesis predicts “something may change”, almost any result can be counted as support.

A strong test requires a more constrained prediction.

Evidence is more informative when the hypothesis took a real risk.

---

111. Theory, law and hypothesis are not ranks on one ladder

A common misconception says:

“A hypothesis becomes a theory, then a law.”

This is not how the concepts relate.

A hypothesis is a testable candidate claim.

A theory is a broad explanatory framework supported by substantial evidence.

A law describes a regularity, often mathematically.

A theory does not “graduate” into a law.

112. A theory can generate hypotheses

A broad theory contains structures and mechanisms from which narrower hypotheses can be derived.

Testing those hypotheses provides evidence that can support, refine or challenge the theory.

113. A law can coexist with a theory

A law may describe what happens.

A theory may explain why or under what mechanism it happens.

They perform different scientific jobs.

114. Models sit alongside hypotheses and theories

A model is a representation used to reason, explain or predict.

A model can embody a hypothesis.

A hypothesis can be tested using a model.

A theory can contain multiple models.

Do not treat the terms as synonyms.

115. Hypotheses can be local even when theories are broad

A hypothesis may test one narrow consequence of a theory.

If the hypothesis fails, the theory may need revision—but only after checking auxiliary assumptions and alternative explanations.

116. Multiple hypotheses can live within one theory

A theory can generate many predictions across domains.

Some hypotheses may be strongly supported while others reveal limits or missing mechanisms.

Theory development is often piecemeal rather than all-or-nothing.

---

117. Confirmation bias

People naturally notice evidence that supports what they already believe.

Scientists are not immune.

Protection strategies include:

• generate alternatives before testing;
• blind analysis where possible;
• predefine outcomes;
• use independent replication;
• ask what evidence would change your mind.

118. Disconfirmation strategy

Instead of asking only:

“What evidence supports my hypothesis?”

ask:

“What observation would be difficult for my hypothesis to explain?”

Then look for that observation.

119. Motivated reasoning

Scientists may have professional, financial or ideological incentives.

Bias control should therefore be structural, not dependent on perfect personal neutrality.

Useful structures include:

• blinding;
• independent analysis;
• data sharing;
• preregistration;
• peer review;
• replication.

120. Blinding

Blinding can reduce bias when expectations could influence:

• measurement;
• treatment;
• outcome assessment;
• analysis.

Not all studies can be blinded, but where possible it strengthens hypothesis testing.

121. Randomisation

Randomisation helps make treatment groups comparable on average with respect to both known and unknown factors.

It is a powerful tool for causal hypotheses.

Randomisation does not guarantee perfect balance in one finite sample, but it creates a principled basis for inference.

122. Controls

A control creates a reference condition.

Without an appropriate comparison, it can be difficult to know what the outcome would have been without the tested factor.

Controls are therefore tools for alternative-hypothesis management.

123. Placebo effects and expectation

In human studies, participants’ expectations can influence outcomes.

Placebo controls can help distinguish treatment-specific effects from expectation and context.

124. Observer effects

The act of measurement can influence the system.

In social research, participants may change behaviour when observed.

In physical systems, measurement can also disturb conditions.

The measurement process may therefore need to become part of the hypothesis model.

125. Confounding

A confounder is associated with both the proposed cause and the outcome in a way that can create or distort an association.

Hypothesis testing should ask:

• what variables could explain both?
• can they be controlled by design?
• can they be adjusted statistically?
• could unmeasured confounding remain?

126. Mediation

A mediator lies on a proposed causal pathway.

If A affects M and M affects B, then M may mediate the effect of A on B.

Testing mediation can strengthen a mechanistic hypothesis, but requires careful causal assumptions.

127. Moderation

A moderator changes the strength or direction of a relationship.

Example:

A treatment works differently at different temperatures, ages or baseline states.

Moderators help define hypothesis boundaries.

128. Interaction effects

Two factors can combine non-additively.

The effect of A may depend on B.

Hypotheses that assume simple independent effects can fail if interactions are important.

129. Dose-response hypotheses

A dose-response pattern can strengthen causal reasoning when the mechanism predicts it.

But not every system is monotonic.

Some show:

• thresholds;
• saturation;
• U-shapes;
• inverted U-shapes.

The shape should come from theory or prior evidence, not be invented after seeing the data.

130. Time-order matters for causality

A cause must precede its effect.

Cross-sectional associations can be ambiguous because temporal order is unclear.

Longitudinal data can help distinguish directions, though confounding may remain.

---

131. Hypothesis stress test

Before a study begins, challenge the hypothesis:

• What if the effect is smaller?
• What if the relationship is non-linear?
• What if the effect reverses outside the tested range?
• What if the measure is biased?
• What if the comparison group differs?
• What if the mechanism is wrong but the prediction still holds?
• What if another hypothesis predicts the same result?

A hypothesis that survives conceptual stress testing is easier to test cleanly.

132. Hypothesis revision

Revision should be specific.

Weak revision:

“The hypothesis was sort of right.”

Stronger revision:

“The relationship appears only above Threshold T and is absent below it; the original hypothesis should be restricted to that range.”

Revision should make new predictions.

133. Hypothesis retirement

A hypothesis may deserve retirement when:

• multiple strong tests fail;
• alternative hypotheses consistently predict better;
• support depends on ad hoc exceptions;
• the proposed mechanism conflicts with robust evidence;
• replications repeatedly fail across valid domains.

Retirement is not defeat.

It is scientific learning.

134. Hypothesis resurrection requires new evidence

An old rejected hypothesis can return if:

• new evidence emerges;
• measurement improves;
• a hidden boundary is discovered;
• a revised mechanism makes new successful predictions.

Science does not ban ideas forever; it requires them to earn reconsideration.

135. Competing models should make commitments

If two models can be tuned to fit every existing observation, ask for predictions in a new domain.

Scientific progress often comes from moving into conditions where the models diverge.

136. Evidence should update a hypothesis, not merely decorate it

A hypothesis is not a slogan to which evidence is attached afterward.

Evidence should have the power to:

• increase confidence;
• decrease confidence;
• change scope;
• change mechanism;
• force retirement.

If no evidence can ever change the claim, the claim is not functioning scientifically.

---

137. Hypothesis Design Laboratory

The next sections turn hypothesis thinking into a reusable design process. Each laboratory begins with a vague idea and forces it toward a claim that could actually lose.

Lab 1 — “Plants grow better with more light”

Problems:

• What does “better” mean?
• How much more light?
• Over what time?
• For which plants?
• Within what range?

Refined hypothesis:

“For seedlings of Species S grown for 21 days under otherwise comparable conditions, increasing daily light exposure from L1 to L2 will increase mean dry mass.”

Derived prediction:

The L2 group will show greater mean dry mass at Day 21.

Competing hypothesis:

The difference is caused by temperature associated with the lighting setup rather than light itself.

Design consequence:

temperature must be monitored or controlled.

Lab 2 — “Music improves concentration”

Vague terms:

• music;
• improves;
• concentration.

Operational choices might include:

• music type and volume;
• task duration;
• accuracy;
• reaction time;
• off-task behaviour.

The exercise teaches that operationalisation is part of hypothesis formation, not clerical detail added later.

Lab 3 — “Exercise reduces stress”

Alternative hypotheses:

• exercise causes stress reduction;
• people with lower stress exercise more;
• sleep quality causes both;
• social support associated with exercise explains the pattern.

A randomised intervention can strengthen causal inference, but ethical and practical constraints shape the design.

Lab 4 — “Students learn better in small groups”

What is “small”?

What is “learn better”?

What content?

What teaching quality?

What prior attainment?

A meaningful hypothesis might compare transfer performance after the same instructional content delivered under controlled group-size conditions.

The hypothesis becomes stronger when the educational mechanism is specified, such as increased feedback or more opportunities to explain.

Lab 5 — “More practice always improves performance”

The word “always” creates an unnecessary universal claim.

Possible boundaries:

• fatigue;
• incorrect practice;
• diminishing returns;
• overtraining;
• interference.

Better hypothesis:

“Within the tested practice range and with corrective feedback, additional retrieval trials improve delayed accuracy up to a point.”

Lab 6 — “This medicine works”

Questions:

• for which condition?
• which population?
• what dose?
• what outcome?
• compared with what?
• over what time?

Medical hypotheses require both benefit and harm outcomes to avoid a one-sided scientific picture.

Lab 7 — “The new sensor is better”

“Better” can mean:

• more accurate;
• more precise;
• more sensitive;
• faster;
• cheaper;
• more robust.

The hypothesis must identify the performance dimension.

Lab 8 — “The algorithm is fairer”

Fairness has multiple definitions.

A testable hypothesis must specify:

• which fairness criterion;
• which groups;
• which decision threshold;
• which outcome distribution;
• which trade-offs matter.

This shows how normative language must be translated carefully before statistical testing.

Lab 9 — “The species is declining because of habitat loss”

Competing hypotheses:

• habitat loss;
• disease;
• predation;
• climate;
• harvesting;
• survey detectability changes.

Different hypotheses predict different spatial and temporal patterns.

Lab 10 — “The machine overheats because the fan is failing”

Predictions:

• fan speed should be lower;
• temperature should improve after fan replacement;
• other cooling components should be normal.

Alternative hypotheses can be tested through diagnostic substitution.

Lab 11 — “The exam score dropped because the student forgot the topic”

Competing explanations:

• retrieval failure;
• question-reading failure;
• timing;
• transfer;
• fatigue;
• unfamiliar representation.

A tutor can design diagnostic questions that isolate each possibility.

Lab 12 — “The climate model is wrong because one year was cooler”

First ask what the model predicted:

• annual exact temperature?
• long-term trend?
• distribution?
• conditional projection?

A single observation may or may not be a relevant falsification depending on the hypothesis target.

Lab 13 — “This correlation proves causation”

Replace the claim with explicit causal alternatives.

Ask:

• Could B cause A?
• Could C cause both?
• Could selection create the association?
• Could measurement create the pattern?

Causal hypotheses require stronger evidence than predictive association.

Lab 14 — “The effect is zero because p > .05”

Rewrite the question:

“Which effect sizes remain compatible with the data?”

Use confidence intervals, power or equivalence testing.

A non-significant test does not automatically support exact zero.

Lab 15 — “The effect is real because p < .05”

Ask:

• What is the effect size?
• What is the uncertainty?
• Was the hypothesis pre-specified?
• How many tests were tried?
• Will it replicate?

Significance is not the end of inference.

Lab 16 — “The treatment works because every treated participant improved”

Without a control, improvement may reflect:

• natural recovery;
• regression to the mean;
• expectation;
• concurrent treatment;
• measurement learning.

Counterfactual reasoning is essential.

Lab 17 — “The theory predicted the result”

Ask whether:

• the prediction was written before data;
• the result was one of many possible predictions;
• the theory predicted a narrow outcome;
• alternative theories predicted the same result.

A prediction’s evidential value depends on risk and timing.

Lab 18 — “The replication failed, so the original study was fraud”

Too strong.

Possible causes include:

• sampling variation;
• protocol difference;
• context dependence;
• measurement;
• false positive;
• effect heterogeneity.

Fraud is one possible explanation only when additional evidence supports it.

Lab 19 — “The model failed one case, so it is useless”

If the hypothesis or model is probabilistic, individual errors are expected.

Evaluate repeated calibration and comparative skill.

Lab 20 — “The hypothesis survived, so it is true”

Survival increases support under the tested conditions.

It does not eliminate every untested alternative.

Future severe tests remain relevant.

---

138. Hypothesis Failure Atlas

Failure 1 — Untestable language

Pattern: key terms cannot be connected to observations.

Repair: operationalise the claim or recognise that it lies outside empirical testing.

Failure 2 — No possible disconfirming evidence

Pattern: every outcome is counted as support.

Repair: state what observation would make the hypothesis less credible.

Failure 3 — Universal claim from narrow data

Pattern: “always”, “all”, or “never” based on a limited sample.

Repair: define the tested domain and boundary.

Failure 4 — Hypothesis written after outcome

Pattern: exploratory discovery is presented as prior prediction.

Repair: label it exploratory and test on new data.

Failure 5 — Only one hypothesis considered

Pattern: every observation is interpreted through the preferred idea.

Repair: generate plausible alternatives before testing.

Failure 6 — Alternatives predict the same result

Pattern: the planned test cannot discriminate.

Repair: find a condition where predictions diverge.

Failure 7 — Poor operational definition

Pattern: measured variable is a weak proxy for the concept.

Repair: improve measurement or triangulate with another measure.

Failure 8 — Measurement ceiling or floor

Pattern: instrument cannot detect expected differences at high or low values.

Repair: choose a measure with suitable range.

Failure 9 — Low power

Pattern: study is unlikely to detect scientifically relevant effects.

Repair: improve sample size, measurement or design.

Failure 10 — Overpowered trivial effect

Pattern: huge sample makes tiny difference statistically significant.

Repair: focus on effect size and practical meaning.

Failure 11 — Confounding

Pattern: another variable can explain the association.

Repair: improve design, randomise, match, stratify or model confounders.

Failure 12 — Reverse causation

Pattern: outcome may influence predictor.

Repair: use temporal data or causal design.

Failure 13 — Selection bias

Pattern: sample inclusion depends on factors related to exposure and outcome.

Repair: redesign sampling or model selection process.

Failure 14 — Attrition bias

Pattern: dropout differs across groups or outcomes.

Repair: examine missingness mechanism and sensitivity.

Failure 15 — Outcome switching

Pattern: primary outcome changes after results are seen.

Repair: preregister, disclose changes, label exploratory outcomes.

Failure 16 — Multiple testing

Pattern: many hypotheses tested with no correction or transparency.

Repair: control error rates and replicate discoveries.

Failure 17 — Optional stopping

Pattern: data collection stops when significance appears.

Repair: predefine stopping rules or use valid sequential methods.

Failure 18 — Post-hoc subgroup

Pattern: effect appears only in one subgroup found after searching many.

Repair: label exploratory and test prospectively.

Failure 19 — Model misspecification

Pattern: statistical test assumptions do not match the data structure.

Repair: diagnose residuals and use appropriate model.

Failure 20 — Non-independence ignored

Pattern: repeated or clustered observations treated as independent.

Repair: use hierarchical, repeated-measures or cluster-aware methods.

Failure 21 — Pseudoreplication

Pattern: many measurements from one experimental unit are treated as many independent units.

Repair: identify true unit of replication.

Failure 22 — Inadequate control

Pattern: comparison does not estimate the relevant counterfactual.

Repair: redesign reference condition.

Failure 23 — Treatment contamination

Pattern: control group receives part of treatment or vice versa.

Repair: improve separation and monitor adherence.

Failure 24 — Expectation bias

Pattern: participants or observers behave differently because they know condition.

Repair: blind where possible.

Failure 25 — Instrument drift

Pattern: measurement changes over time.

Repair: recalibrate and audit measurement.

Failure 26 — Hypothesis overfitting

Pattern: hypothesis becomes increasingly tailored to one dataset.

Repair: test revised hypothesis on independent data.

Failure 27 — Ad hoc exception

Pattern: every failure generates an exception that makes no new prediction.

Repair: require independent rationale and prospective test.

Failure 28 — Confirmation-only sampling

Pattern: scientist looks only where expected positive cases occur.

Repair: sample comparison cases and disconfirming conditions.

Failure 29 — Survivorship bias

Pattern: analysis includes only successful or surviving cases.

Repair: account for missing failures.

Failure 30 — Publication bias

Pattern: positive tests dominate the literature.

Repair: study registration, null-result publication, meta-analysis correction.

Failure 31 — Poor replication fidelity

Pattern: “replication” changes crucial procedures unknowingly.

Repair: document protocol and distinguish direct from conceptual replication.

Failure 32 — Context dependence mistaken for universal failure

Pattern: replication fails in a different domain.

Repair: test moderator hypothesis.

Failure 33 — One replication treated as final

Pattern: a single repeat overrides the full evidence base.

Repair: synthesise multiple studies and quality differences.

Failure 34 — Statistical null confused with substantive null

Pattern: failure to reject is treated as proof of no scientifically meaningful effect.

Repair: use interval estimates, power and equivalence tests.

Failure 35 — Mechanism claimed from association

Pattern: correlational data used to assert causal pathway.

Repair: collect mechanistic or intervention evidence.

Failure 36 — Theories treated as guesses

Pattern: “just a theory” implies lack of evidence.

Repair: distinguish theory from hypothesis and law.

Failure 37 — Hypothesis treated as an early-stage theory

Pattern: linear hierarchy misconception.

Repair: teach distinct scientific roles.

Failure 38 — Prior ignored

Pattern: extraordinary claim and ordinary claim treated as equally plausible before evidence.

Repair: consider base rates and prior evidence.

Failure 39 — Prior dominates despite strong evidence

Pattern: old belief never updates.

Repair: allow sufficiently diagnostic evidence to move beliefs.

Failure 40 — Missing hypothesis

Pattern: analysis tests only available variables, not a coherent scientific claim.

Repair: return to question and model before statistics.

---

139. The hypothesis tree

A hypothesis can be organised as a tree.

Root question: Why does Outcome Y differ?

Branch A: Cause X1.

Branch B: Cause X2.

Branch C: measurement artifact.

Branch D: selection or confounding.

Each branch generates predictions.

Testing should prune branches whose predictions fail and split surviving branches into more precise hypotheses.

140. Build the tree before choosing the test

Statistics should come after the scientific structure.

If the question is:

“Which of these mechanisms explains the pattern?”

then a simple difference test may be insufficient.

The hypothesis tree helps identify what evidence would discriminate mechanisms.

141. Alternative-hypothesis matrix

Hypothesis	Prediction 1	Prediction 2	Distinctive evidence
A	Outcome rises	Marker M rises	M changes before outcome
B	Outcome rises	Marker M unchanged	different pathway
C	apparent rise only	measurement shifts	independent instrument disagrees

The matrix makes clear why one observation can be weak evidence if all hypotheses predict it.

142. Prediction uniqueness

A prediction is most discriminating when few competing hypotheses also predict it.

This is why surprising successful predictions can be powerful.

143. Prediction precision

Narrow predictions are more informative if justified.

But unjustified precision can create brittle tests and false confidence.

The goal is justified specificity.

144. Prediction independence

Evidence is stronger when the prediction was not built from the same data later used to test it.

Hold-out tests, preregistration and independent replication create cleaner separation.

145. Hypothesis entropy intuition

Before an experiment, several hypotheses may remain plausible.

A good experiment reduces uncertainty over that hypothesis set.

Tests that produce the same expected result under every hypothesis reduce little uncertainty.

This is an information-theoretic way to think about experiment value.

146. Expected information gain

If one measurement would sharply distinguish alternatives, it may be more valuable than collecting many redundant measurements.

Experimental design can therefore ask:

Which observation would teach us the most about which hypothesis is right?

147. Active learning in science

Scientists often choose the next experiment based on current uncertainty.

This creates an iterative loop:

hypotheses → choose informative test → observe → update → choose next test

Science is adaptive, not always a fixed one-shot plan.

148. Hypothesis portfolios

In complex systems, it can be useful to maintain several live hypotheses rather than committing too early.

Each new observation updates the portfolio.

This reduces premature closure.

149. Premature closure

Premature closure occurs when researchers stop considering alternatives too early.

It is common in:

• medical diagnosis;
• engineering troubleshooting;
• scientific model selection;
• educational diagnosis.

A hypothesis portfolio keeps alternatives visible until evidence discriminates them.

150. Occam’s razor

When two hypotheses explain the evidence equally well, scientists often prefer the simpler one.

But “simpler” needs care.

It can mean:

• fewer assumptions;
• fewer free parameters;
• less ad hoc complexity;
• more compression;
• greater generality.

Simplicity is a guide, not proof.

151. Complexity can be justified

A more complex hypothesis can be better if it:

• predicts new data better;
• explains anomalies;
• unifies phenomena;
• has independent mechanistic support.

Occam’s razor does not say “always choose the fewest words”.

152. Parsimony and overfitting

Parsimony protects against adding one new parameter for every anomaly.

But excessive simplicity can underfit real structure.

Model selection balances fit and complexity.

153. AIC and BIC intuition

Information criteria such as AIC and BIC combine goodness of fit with penalties for complexity.

They are tools for comparing statistical models, not universal truth machines.

Results depend on model set and assumptions.

154. Cross-validation as hypothesis stress

Cross-validation asks whether a model or hypothesis-derived predictor generalises beyond the data used to fit it.

Repeated predictive success can increase confidence that the relationship is stable.

155. Mechanistic evidence can rescue weak prediction—and vice versa

A mechanism may be compelling but predictions may be quantitatively weak because parameters are uncertain.

A predictive model may be strong but mechanism unclear.

Scientific confidence improves when mechanism and prediction converge.

156. Explanatory power

A good hypothesis can explain multiple observations with one mechanism.

This unification is scientifically valuable.

But explanation should still create testable consequences.

157. Scope

A broad hypothesis explains more phenomena.

A narrow hypothesis can be more precise.

Science often develops by moving between them:

• narrow test;
• broader generalisation;
• boundary testing;
• refined theory.

158. Fruitfulness

A fruitful hypothesis generates new questions, predictions and experiments.

Even before final confirmation, a hypothesis can be scientifically valuable if it opens a productive research programme.

159. Coherence

A hypothesis should fit with well-established knowledge unless it provides strong evidence for overturning that knowledge.

Extraordinary departures require correspondingly strong tests.

160. Novelty is not enough

A surprising hypothesis can attract attention.

But scientific merit comes from testability, evidence, explanatory value and predictive success—not surprise alone.

---

161. The Complete Hypothesis Protocol

A hypothesis-driven investigation can be organised around twelve questions.

1. Question: What phenomenon needs explanation or prediction?
2. Candidate hypotheses: What plausible ideas could account for it?
3. Domain: Under what conditions should each hypothesis apply?
4. Operationalisation: What exactly will be measured?
5. Predictions: What should we observe if each hypothesis is correct?
6. Discrimination: Where do the hypotheses predict different outcomes?
7. Design: What test can expose those differences?
8. Controls: What alternative explanations must be reduced?
9. Analysis: How will evidence be evaluated?
10. Failure criterion: What result would count against the hypothesis?
11. Update: How will support change after the evidence?
12. Next test: What new question follows?

This is not a rigid sequence. Science often loops backward. A measurement problem can force a new operational definition. An unexpected result can generate a new hypothesis. A failed replication can reveal a hidden moderator.

162. Step 1 — Begin with a real question

Weak:

“I need a hypothesis for my project.”

Stronger:

“Why did Outcome Y change under Condition C?”

Hypotheses should grow from questions, not from the need to fill a template.

163. Step 2 — Generate a minimum of two plausible alternatives

Even when one explanation seems obvious, deliberately generate another.

This forces the design to distinguish rather than merely confirm.

164. Step 3 — State the domain

Ask:

• Which population?
• Which range?
• Which environment?
• Which time scale?
• Which measurement system?

A hypothesis without a domain tends to become over-broad.

165. Step 4 — Operationalise key terms

Translate abstract terms into observable variables.

Examples:

• “better learning” → delayed transfer score;
• “plant health” → defined physiological or growth measures;
• “machine reliability” → failure-free operating hours;
• “stress” → specified self-report or physiological measure.

166. Step 5 — Derive predictions

Write what should happen before collecting the key data.

Prediction quality improves when it includes:

• direction;
• comparison;
• time window;
• domain;
• magnitude when justified.

167. Step 6 — Find the disagreement

If every hypothesis predicts the same outcome, the test is weak.

Ask:

“What observation would favour A over B?”

168. Step 7 — Design the test around that disagreement

The best experiment may not be the one that maximises the signal.

It may be the one that maximises information about which hypothesis is correct.

169. Step 8 — Protect the comparison

Controls, randomisation, blinding and standardisation are ways to reduce alternative explanations.

Use only the tools appropriate to the system and ethical constraints.

170. Step 9 — Predefine analysis where confirmation matters

For confirmatory research, specify:

• primary outcome;
• statistical model;
• exclusion rules;
• stopping rule;
• decision criterion.

Exploratory analysis can follow, but should remain clearly labelled.

171. Step 10 — Define failure before the outcome

Ask:

“What result would make us reduce confidence in this hypothesis?”

If the answer is “none”, the hypothesis is protected rather than tested.

172. Step 11 — Update proportionally

One study should not produce infinite certainty.

Update based on:

• test quality;
• effect magnitude;
• measurement;
• alternative explanations;
• replication;
• prior evidence.

173. Step 12 — Generate the next hypothesis

Scientific inquiry is recursive.

A supported result can ask:

• what mechanism?
• which boundary?
• which population?
• what dose?
• what long-term effect?

A failed result can ask:

• wrong hypothesis?
• wrong measure?
• wrong domain?
• missing moderator?
• insufficient power?

---

174. Hypothesis Practice Bank: 30 reusable prompts

Practice 1 — Turn a vague claim into a hypothesis

Start with “sleep affects learning”. Ask the student to specify population, sleep variable, learning measure and time window.

Practice 2 — Name two alternatives

Observation: students who attend tutoring score higher. Generate at least two explanations besides tutoring causing improvement.

Practice 3 — Derive a prediction

Hypothesis: increased temperature speeds Process P within Range R. Write one measurable prediction.

Practice 4 — State a disconfirming result

Ask what observation would count against the hypothesis, not merely fail to support it weakly.

Practice 5 — Operationalise “motivation”

Propose two different measures and discuss how they might capture different parts of the concept.

Practice 6 — Operationalise “plant health”

Choose height, leaf count, dry mass or another measure. Discuss why the choice matters.

Practice 7 — Find the domain

Hypothesis: “more water increases growth”. Ask where that statement is likely to stop being true.

Practice 8 — Find the confounder

Observation: children with more books at home have higher reading scores. Generate possible third variables.

Practice 9 — Reverse causation

Observation: happier people socialise more. Ask whether happiness causes socialising, socialising causes happiness, or both.

Practice 10 — Design a discriminating experiment

Hypothesis A says heat causes the change. Hypothesis B says light causes the change. Design setups where heat and light differ independently.

Practice 11 — Separate scientific and statistical hypothesis

Write the substantive hypothesis about a treatment, then the null model used for analysis.

Practice 12 — Reject is not prove

Given p = .01, list five things that are still not established automatically.

Practice 13 — Non-significant is not zero

Given p = .20 and a wide confidence interval, ask what effect sizes remain compatible.

Practice 14 — Hypothesis after result

A surprising subgroup effect appears. Decide what can be claimed now and what needs future testing.

Practice 15 — Replication question

A replication fails. Generate at least five hypotheses explaining the disagreement.

Practice 16 — Boundary test

A relationship holds from 10–30°C. Design tests near and beyond the boundary rather than only at 20°C.

Practice 17 — Prediction uniqueness

Three hypotheses all predict Outcome X. Ask what second measurement could separate them.

Practice 18 — Auxiliary assumption

A prediction fails. List one main hypothesis and three auxiliary assumptions that might have failed.

Practice 19 — Ad hoc rescue

Given a failed hypothesis, write one legitimate revision and one ad hoc excuse. Explain the difference.

Practice 20 — Competing causal diagrams

Represent A→B, B→A and C→A/B. Discuss which data could distinguish them.

Practice 21 — Prior plausibility

Compare a claim consistent with established physics and one requiring a major revision. Discuss why evidence thresholds may differ.

Practice 22 — Bayesian update

Start with two hypotheses and prior odds. Reveal evidence with a likelihood ratio. Update qualitatively or quantitatively.

Practice 23 — Publication bias

Imagine ten studies, only one significant. Show how publishing only one distorts the apparent evidence.

Practice 24 — Multiple testing

Test 20 independent null hypotheses at α=.05. Discuss expected false positives and why replication matters.

Practice 25 — Pseudoreplication

Measure 100 leaves from one plant and ask whether the sample size for independent plants is 100 or 1.

Practice 26 — Randomisation logic

Explain what random assignment protects against and what it does not automatically solve.

Practice 27 — Blinding logic

Identify where knowledge of condition could bias measurement or behaviour.

Practice 28 — Equivalence question

Instead of asking “Is there a difference?”, define a scientifically meaningful equivalence margin.

Practice 29 — Hypothesis retirement

List what repeated evidence would justify abandoning a preferred idea.

Practice 30 — Full hypothesis cycle

Take one question through alternatives, operationalisation, predictions, experiment, analysis, evidence, update and next test.

---

175. Decision Framework: Which kind of hypothesis are you testing?

Hypothesis type	Main question	Typical evidence
Descriptive	What pattern exists?	observation, survey, measurement
Comparative	Do groups differ?	controlled comparison, sampling
Associational	Do variables covary?	correlation, regression, longitudinal data
Causal	Would changing X change Y?	randomisation, natural experiment, causal design
Mechanistic	Through what process?	intervention on mediator, direct measurement
Predictive	What unseen outcome follows?	out-of-sample forecasting
Boundary	Where does the relationship stop?	range tests, stress tests
Equivalence	Are differences smaller than a meaningful margin?	equivalence test, interval estimates

176. Descriptive hypotheses

Descriptive hypotheses can be foundational.

Example:

“Species richness decreases with elevation in this region.”

No cause is asserted.

The hypothesis can still be tested through systematic sampling.

177. Comparative hypotheses

Comparative hypotheses ask whether two or more conditions differ.

They require careful definition of groups and comparable measurement.

178. Associational hypotheses

Associational hypotheses ask whether variables change together.

They can be useful for prediction and theory generation.

They should not automatically be written in causal language.

179. Causal hypotheses

Causal hypotheses ask what would happen under intervention.

This counterfactual question requires stronger assumptions and design.

180. Mechanistic hypotheses

Mechanistic hypotheses propose the pathway connecting cause to outcome.

They often predict intermediate measurements.

These intermediate predictions can distinguish mechanisms that produce the same final outcome.

181. Predictive hypotheses

Predictive hypotheses commit to unseen outcomes.

They are strongest when evaluated prospectively or out of sample.

182. Boundary hypotheses

Boundary hypotheses identify where a relationship changes, saturates, reverses or disappears.

These are essential for preventing overgeneralisation.

183. Equivalence hypotheses

Equivalence hypotheses define a region of differences considered scientifically negligible.

They are useful when the scientific question concerns similarity rather than difference.

---

184. Hypothesis Writing Workshop

Weak: “Water affects plants.”

Why weak: direction, amount, outcome, domain and comparison are missing.

Improved: “Within the tested range, increasing daily water supply will increase mean dry mass after 21 days in Species S under comparable light and soil conditions.”

Now the hypothesis:

• defines range;
• defines outcome;
• defines population;
• defines controls;
• predicts direction.

Weak: “Students using Method A will learn better.”

Improved:

“Students assigned to retrieval practice for equal total study time will score higher on a delayed transfer test one week later than students assigned to rereading.”

The improved version operationalises learning as delayed transfer.

Weak: “The machine fails because it is old.”

Improved:

“Components with operating age above Threshold T will show a higher hazard of Failure Mode F after controlling for measured load and temperature.”

This is testable, but the causal interpretation still depends on design.

Weak: “Pollution causes asthma.”

Improved research hypothesis:

“Higher long-term exposure to Pollutant P is associated with increased incidence of asthma symptoms in Population Q after accounting for measured confounders.”

If causality is intended, stronger causal design and language are needed.

Weak: “The model is accurate.”

Improved predictive hypothesis:

“On data collected after the training period, Model M will achieve lower mean absolute error than Baseline B over Forecast Horizon H.”

Now performance has a target, baseline and horizon.

---

185. Hypothesis Evaluation Rubric

Dimension	Weak	Developing	Strong
Testability	no clear evidence could bear on claim	some observable consequence	clear possible support and failure
Specificity	vague terms	some variables defined	target, domain and relation explicit
Alternatives	none considered	one alternative	plausible competing hypotheses mapped
Prediction	after-the-fact	direction only	pre-specified discriminating prediction
Measurement	construct undefined	single rough proxy	valid operationalisation with limitations stated
Scope	universal claim	some boundary	domain and boundary explicit
Updateability	claim never changes	revision possible	clear evidence-dependent update rule

186. Why this rubric matters

Students are often graded on whether a hypothesis “sounds scientific”.

A stronger rubric evaluates whether the hypothesis creates real contact with evidence.

187. The deepest hypothesis habit

Before defending an idea, ask:

What would I expect to see if I were wrong?

This question protects science from becoming an exercise in proving what we already believe.

---

188. Advanced Hypothesis Casebook: from observation to discriminating test

The following cases move beyond the introductory examples. Each case asks the same scientific question in a different domain: What are the plausible hypotheses, what does each predict, and what observation would help us choose?

Case 16 — A reaction slows unexpectedly

A chemical reaction initially follows the expected rate relationship, then slows sharply.

Candidate hypotheses:

• a reactant is being depleted;
• a product inhibits the reaction;
• temperature fell;
• the catalyst lost activity;
• the measuring instrument saturated.

A strong next step is not “repeat exactly the same thing and hope”. Instead, derive distinct predictions. If reactant depletion is responsible, adding more reactant should restore rate under otherwise comparable conditions. If catalyst deactivation is responsible, fresh catalyst may restore rate. If the instrument saturated, an independent measurement method should disagree with the original sensor.

This case shows why good hypotheses create different intervention predictions.

Case 17 — A metal corrodes faster near the coast

Observation:

Outdoor metal structures near the coast corrode faster than similar structures inland.

Hypotheses might involve:

• salt exposure;
• humidity;
• temperature;
• different maintenance practices;
• different material batches.

A useful design could compare matched material samples under controlled salt and humidity conditions. A field study could also measure actual exposure. The point is to separate a geographic label—“coastal”—into mechanisms that can be tested.

Case 18 — One lake suddenly loses oxygen

Possible explanations include:

• nutrient-driven algal growth followed by decomposition;
• thermal stratification;
• reduced inflow;
• pollutant discharge;
• sensor malfunction.

Each predicts a different pattern across depth, time, nutrient concentration and independent oxygen measurements. The hypothesis test should sample those dimensions rather than rely on one surface reading.

Case 19 — An animal population moves uphill

A long-term survey finds that a species is observed at higher elevation over time.

Hypotheses:

• temperature change shifts suitable habitat;
• food resources shift uphill;
• low-elevation habitat is lost;
• survey effort changes;
• predators or competitors change distribution.

Testing requires more than plotting elevation against year. Researchers should measure the proposed drivers and check whether the spatial pattern matches each mechanism.

Case 20 — A star dims periodically

Candidate hypotheses:

• a transiting planet;
• starspots;
• binary stellar motion;
• instrumental artifact.

A planetary-transit hypothesis predicts repeatable timing and a characteristic wavelength-independent or weakly wavelength-dependent pattern under simple conditions. Stellar activity may produce different spectral or temporal signatures. Independent instruments help test the artifact hypothesis.

The important lesson is that “a dip happened” supports several hypotheses. Follow-up evidence earns discrimination.

Case 21 — A galaxy rotates faster than visible matter predicts

Historically, observations of galactic rotation helped motivate hypotheses about unseen mass distributions, alternative gravity models and measurement/systematic explanations.

A strong scientific programme asks how each candidate explains not only one rotation curve but additional phenomena. Hypotheses become more credible when they make successful predictions beyond the observation that originally motivated them.

Case 22 — A fossil appears outside the expected time range

Possible explanations:

• the geological layer was dated incorrectly;
• the fossil was reworked from older sediment;
• the species’ known time range is incomplete;
• the specimen was misidentified.

Independent dating, sedimentology and morphological analysis can distinguish the alternatives. Scientific revision should target the assumption that failed rather than immediately overturning the entire chronology.

Case 23 — A seismic signal appears before several eruptions

Hypothesis:

The signal reflects magma movement and therefore raises eruption probability.

Alternative:

The signal reflects unrest that often occurs without eruption.

The right question is not whether the signal ever appears before eruptions. It is how often eruptions follow the signal, how often the signal occurs without eruption, and whether other measurements add discrimination.

Prediction requires both sensitivity and false-alarm analysis.

Case 24 — A new material appears stronger in the laboratory

Hypotheses:

• the material genuinely has greater tensile strength;
• sample geometry differs;
• testing machine calibration differs;
• manufacturing defects differ between batches.

A valid mechanical hypothesis requires standardised specimen geometry, calibration and enough independent samples. One spectacular specimen is not the population.

Case 25 — A bridge vibrates unexpectedly

Candidate mechanisms:

• wind excitation;
• traffic loading;
• structural resonance;
• foundation motion;
• sensor artifact.

Mechanistic hypotheses should predict frequency content, timing and response to interventions. If changing traffic has no effect but vibration tracks wind speed near a structural natural frequency, support shifts.

Case 26 — A battery degrades faster than expected

Hypotheses:

• high temperature accelerates degradation;
• charge protocol is damaging;
• manufacturing lot differs;
• measurement of capacity has drifted.

A factorial design can separate temperature and charge-protocol effects. Lot should be included or randomised. Independent calibration tests the measurement hypothesis.

Case 27 — A medical test shows more positive cases

Possible hypotheses:

• true prevalence increased;
• test sensitivity increased;
• testing expanded into a different population;
• classification threshold changed;
• data coding changed.

Observed positive counts alone cannot identify which explanation is correct. Rates, denominators and measurement changes matter.

Case 28 — Hospital mortality falls after a new protocol

Hypothesis:

The protocol caused improved survival.

Alternatives:

• patient mix changed;
• another treatment changed simultaneously;
• coding changed;
• seasonality;
• regression to the mean.

A controlled design, interrupted time-series analysis or matched comparison may help. Causal hypotheses require a credible counterfactual.

Case 29 — A vaccine appears less effective in one region

Possible modifiers:

• variant distribution;
• time since vaccination;
• age distribution;
• prior infection;
• case ascertainment;
• storage or delivery differences.

The correct scientific response is not immediately “the vaccine stopped working”. Generate moderator hypotheses and test them.

Case 30 — A psychological effect appears only in one laboratory

Hypotheses:

• the effect is context-dependent;
• the original estimate was a false positive;
• procedure details differ;
• participant populations differ;
• experimenter expectations matter.

Multi-lab replication can test whether the laboratory itself moderates the effect.

Case 31 — A classroom intervention raises immediate scores but not delayed transfer

Possible interpretations:

• the intervention improves short-term performance but not durable learning;
• the delayed test measures a different construct;
• practice effects inflated the immediate test;
• students did not retain the strategy independently.

This case is why “learning” must be operationalised carefully. Immediate success and durable transfer are different hypotheses.

Case 32 — Small-group tuition seems more effective

Possible causal mechanism:

smaller groups permit more diagnostic feedback and student explanation.

Alternative explanations:

• families selecting small groups differ;
• tutors differ;
• curriculum differs;
• students begin at different levels.

A useful hypothesis test should measure the mechanism—feedback, prompt count, participation—not only final scores.

Case 33 — A machine-learning model predicts disease accurately

Research hypothesis:

the model contains stable predictive information available before diagnosis.

Competing explanations:

• data leakage;
• site-specific workflow signals;
• post-diagnosis codes;
• sampling bias.

External validation at a new site and temporal holdout testing are discriminating tests.

Case 34 — An AI model improves after adding a feature

Hypothesis:

the feature contains genuine predictive signal.

Alternative:

the feature leaks the target or is a site-specific proxy.

Test performance should be evaluated prospectively and across shifted environments. Feature importance alone cannot distinguish signal from leakage.

Case 35 — A recommendation system changes user behaviour

Hypothesis:

recommendations causally increase engagement.

Alternative:

the system recommends items already likely to be chosen.

A randomised A/B test can estimate incremental effect. Observational click-through rates alone cannot establish causality.

Case 36 — A financial strategy backtests well

Hypotheses:

• the strategy captures a persistent market relationship;
• the backtest overfits noise;
• transaction costs erase the effect;
• survivorship bias inflates results;
• future market regime differs.

Out-of-sample and live forward testing are stronger than repeated optimisation on one historical period.

Case 37 — A forecast consistently misses during extremes

Hypothesis:

the model omits a nonlinear mechanism active only in extreme regimes.

Alternative:

measurements are less reliable during extremes.

Compare independent instruments and regime-specific residuals. The error pattern should guide the next hypothesis.

Case 38 — An archaeological site appears older than expected

Possible explanations:

• dating error;
• contamination;
• misinterpreted stratigraphy;
• the cultural chronology is incomplete.

Independent dating methods and stratigraphic analysis can test alternatives.

Case 39 — A language pattern appears across unrelated cultures

Hypotheses may involve:

• shared cognitive constraints;
• historical diffusion;
• sampling bias;
• coding definitions;
• chance.

Comparative data and phylogenetic methods may help separate inheritance from independent convergence.

Case 40 — A conservation intervention appears to work

Population numbers increase after protection begins.

Possible hypotheses:

• protection caused recovery;
• weather improved;
• survey methods changed;
• migration increased;
• another policy acted simultaneously.

Before–after data become stronger when combined with comparison sites and mechanism evidence.

Case 41 — A river becomes clearer after regulation

Hypothesis:

the regulation reduced pollutant discharge.

Alternative explanations include rainfall variation, industrial closure or measurement change.

A hypothesis about regulation should predict changes in the regulated pollutants, upstream/downstream patterns and timing relative to enforcement.

Case 42 — A species evolves resistance

Hypothesis:

exposure to a control agent creates selection favouring resistant variants.

Predictions can involve:

• resistance frequency increasing under exposure;
• genetic or physiological mechanisms;
• reduced fitness cost in treated environments;
• different dynamics without exposure.

Mechanistic and population evidence can converge.

Case 43 — A novel pathogen spreads faster

Possible hypotheses:

• higher transmissibility;
• immune escape;
• founder effects;
• behavioural change;
• reporting changes.

Growth advantage alone does not identify mechanism. Laboratory, epidemiological and genomic evidence can discriminate alternatives.

Case 44 — One teaching method works for some students but not others

Hypothesis:

prior knowledge moderates the method’s effect.

Prediction:

the treatment effect differs systematically by baseline knowledge.

This should be specified and tested with interaction models rather than discovered through uncontrolled subgroup searching after the fact.

Case 45 — A medicine has no average effect but helps a subgroup

Possible interpretations:

• true effect heterogeneity;
• chance subgroup finding;
• biomarker interaction;
• measurement noise.

Independent prospective testing of the subgroup hypothesis is essential.

---

189. Statistical Hypothesis Testing: 25 common myths

Myth 1 — p < .05 means the hypothesis is 95% likely to be true

No. A p-value is calculated under a specified null model. It is not directly the probability that the scientific hypothesis is true.

Myth 2 — p > .05 proves there is no effect

No. The study may be underpowered or imprecise. Inspect intervals and meaningful-effect thresholds.

Myth 3 — statistical significance means practical importance

No. Effect magnitude and context determine importance.

Myth 4 — a tiny p-value proves the mechanism

No. It may reject a statistical null while leaving many causal explanations possible.

Myth 5 — non-significant results should not be published

False. Well-designed null results can rule out large effects, refine theory and reduce publication bias.

Myth 6 — increasing sample size solves every study problem

No. Large samples do not fix bias, confounding, invalid measurement or leakage.

Myth 7 — randomisation guarantees perfectly equal groups

No. It balances factors probabilistically across repetitions. Finite samples can still differ by chance.

Myth 8 — randomisation fixes poor measurement

No. Outcome and treatment measurement still need validity and reliability.

Myth 9 — control groups are always untreated

No. Controls can receive placebo, standard treatment, alternate condition or another reference appropriate to the question.

Myth 10 — every experiment needs a null hypothesis

No. Scientific experiments test substantive hypotheses; statistical nulls are one analysis framework.

Myth 11 — the alternative hypothesis is automatically the scientist’s favourite explanation

No. Statistical alternatives may represent broad classes of parameter values, not one mechanism.

Myth 12 — a hypothesis must be one sentence

No. Complex hypotheses may require formal models, diagrams, equations and explicit boundary conditions.

Myth 13 — hypotheses are guesses made before facts are known

No. Good hypotheses are informed by theory, prior evidence and observation.

Myth 14 — a hypothesis becomes a theory after enough evidence

No. Hypotheses and theories have different roles.

Myth 15 — laws are proven theories

No. Laws describe regularities; theories explain and organise phenomena.

Myth 16 — falsifiability means one bad result destroys a hypothesis instantly

No. Tests rely on auxiliary assumptions and measurement. Failure requires diagnosis.

Myth 17 — repeated confirmation proves universal truth

No. Future conditions can reveal boundaries.

Myth 18 — an unfalsified hypothesis is supported automatically

No. A hypothesis may survive only weak tests or make vague predictions.

Myth 19 — hypotheses should never change

No. Revision in response to evidence is normal science.

Myth 20 — changing a hypothesis is moving the goalposts

Not necessarily. Transparent, prospectively tested revision is legitimate. Quietly rewriting the original claim after results is not.

Myth 21 — exploratory hypotheses are bad science

No. Exploration is essential. The problem is presenting exploratory discoveries as prior predictions.

Myth 22 — replication should produce identical numbers

No. Replications should produce results compatible with the underlying effect, allowing sampling and contextual variation.

Myth 23 — a failed replication means the original study was fraudulent

No. Many scientific explanations are possible; fraud requires separate evidence.

Myth 24 — a successful replication proves the mechanism

No. It strengthens evidence for the reproducibility of the effect under those conditions, not necessarily one causal mechanism.

Myth 25 — hypothesis testing ends when a paper is published

No. Publication begins broader scrutiny, replication and extension.

---

190. Hypothesis Glossary

Alternative hypothesis

A hypothesis considered in contrast to another, often the null in statistical testing or a competing scientific explanation.

Auxiliary assumption

A supporting assumption required to derive a prediction or interpret a test.

Bayes factor

A ratio comparing how well two statistical hypotheses predict observed data after integrating over parameters under their priors.

Boundary condition

A condition defining where a hypothesis or model is expected to apply.

Causal hypothesis

A claim that intervening on one factor would change another under specified conditions.

Conceptual replication

A new test of the same underlying hypothesis using different methods, measures or contexts.

Confidence interval

An interval produced by a method with a defined long-run coverage property under repeated sampling and model assumptions.

Confirmation bias

A tendency to seek or interpret evidence in ways that favour existing beliefs.

Confounder

A variable that can distort a causal association because it is related to both the proposed cause and outcome.

Directional hypothesis

A hypothesis specifying which direction an effect or difference should take.

Direct replication

A new study attempting to reproduce the original procedure closely with new data.

Discriminating test

A test designed where competing hypotheses make different predictions.

Effect size

A quantitative measure of the magnitude of a difference, association or effect.

Equivalence hypothesis

A claim that differences are smaller than a pre-specified scientifically meaningful margin.

Exploratory analysis

Analysis used to discover patterns or generate hypotheses rather than confirm a pre-specified claim.

Falsifiability

The property that a claim could conflict with possible observations and therefore risk being wrong.

HARKing

Presenting a hypothesis generated after seeing the results as though it had been specified beforehand.

Hypothesis

A testable candidate claim, relationship or explanation that generates expectations about evidence.

Likelihood

The probability of observed data under a specified hypothesis or parameter value, viewed as a function of the hypothesis or parameter.

Mechanistic hypothesis

A hypothesis proposing the process through which an effect occurs.

Moderator

A variable that changes the strength or direction of a relationship.

Null hypothesis

A statistical reference hypothesis, often specifying no difference or a particular parameter value.

Operational definition

The specific procedure used to measure or manipulate a scientific concept.

p-value

Under a specified null model, the probability of obtaining data at least as incompatible with the null as the observed data according to the test statistic.

Power

The probability that a statistical test will reject the null for a specified alternative under its assumptions.

Prediction

An expected observation or outcome derived from a hypothesis and auxiliary assumptions.

Preregistration

Time-stamped documentation of hypotheses, methods or analyses before observing the relevant outcomes.

Pseudoreplication

Treating non-independent measurements as if they were independent experimental units.

Registered report

A publication format in which rationale and methods are peer-reviewed before results are known.

Replication

A new empirical test of a scientific result or hypothesis using new observations.

Reproducibility

The ability to obtain consistent computational or analytic results using the same data, code and procedures, depending on field-specific terminology.

Research hypothesis

The substantive scientific claim motivating an investigation.

Severe test

A test that would have had a high chance of exposing a relevant flaw if the hypothesis were wrong.

Statistical significance

A decision criterion based on a statistical test and threshold; it does not by itself indicate scientific importance.

Theory

A broad explanatory framework supported by evidence that organises facts, mechanisms and predictions.

Type I error

Rejecting a null hypothesis when it is true under the model.

Type II error

Failing to reject a null hypothesis when the specified alternative is true.

---

191. Advanced FAQ

1. Does every scientific study need a hypothesis?

No. Descriptive, exploratory and discovery research can be scientifically valuable without a formal prior hypothesis.

2. Is a hypothesis just an educated guess?

That phrase is too weak. A scientific hypothesis is a testable candidate claim informed by existing knowledge and connected to possible evidence.

3. Does testable mean experimentally testable?

No. Astronomy, geology and evolutionary biology often test hypotheses using observations, natural experiments and historical evidence.

4. Can a hypothesis be about the past?

Yes. A hypothesis about past events can generate predictions about traces that should exist now.

5. Can a hypothesis be mathematical?

Yes. Many hypotheses are expressed through equations, probability models or formal constraints.

6. Can a hypothesis be probabilistic?

Yes. Probabilistic hypotheses predict distributions or probabilities rather than deterministic outcomes.

7. Can two hypotheses both be partly right?

Yes. Complex systems often have multiple contributing causes. Tests may estimate their relative importance rather than choose only one.

8. Can a hypothesis be too broad?

Yes. Broad claims can be difficult to test because the domain and prediction are unclear.

9. Can a hypothesis be too narrow?

Yes. A claim tailored to one dataset may lack generality and become overfit.

10. Is simplicity always better?

No. Simplicity is preferred when explanatory and predictive performance are comparable. Extra complexity is justified when it earns real explanatory or predictive value.

11. How many alternatives should I generate?

Enough to cover plausible mechanisms and important artifacts. There is no fixed number.

12. What if I cannot think of an alternative?

Ask what measurement error, selection process, confounder or reverse causal direction could produce the same observation.

13. Can evidence ever prove a hypothesis?

Formal proof belongs to mathematics and logic. Empirical science usually accumulates support while remaining open to revision.

14. What does it mean to “reject” a hypothesis?

It means evidence has reduced its credibility enough under the testing framework that another explanation or model is preferred. The exact meaning depends on context.

15. Should a scientist want their hypothesis to survive?

Scientists can prefer an idea, but good practice designs tests capable of showing the idea is wrong.

16. Is changing a hypothesis after new evidence dishonest?

No. Transparent revision is core science. Dishonesty arises when revisions are hidden and presented as original predictions.

17. What is the difference between revising and moving goalposts?

Revision acknowledges the prior claim, explains why it changed and subjects the new claim to fresh tests. Moving goalposts rewrites success criteria after the fact to avoid failure.

18. What if a hypothesis cannot be tested today?

It may remain a scientific idea if it could become testable with future technology, but current evidential status should be clear.

19. Is falsifiability the only criterion for science?

No. Scientific practice also values explanatory coherence, evidence, reproducibility, predictive success, measurement quality and fruitful research programmes.

20. How do hypotheses become theories?

They do not simply “become” theories. Theories are broader explanatory frameworks that may contain and generate many hypotheses.

21. What is the difference between a hypothesis and a model?

A hypothesis is a claim. A model is a representation. A model can embody or test a hypothesis.

22. What is the difference between a hypothesis and an assumption?

A hypothesis is typically the claim under investigation. Assumptions are background conditions used to derive predictions or interpret evidence.

23. What is the difference between a hypothesis and a prediction?

The hypothesis states the proposed relationship or explanation; the prediction states what should be observed if that hypothesis and its assumptions are correct.

24. What is the difference between a null hypothesis and “nothing happened”?

A null is a formal statistical reference model. It can specify no difference, a parameter value or another condition; it is not simply “nothing happened”.

25. What is the best evidence for a hypothesis?

Evidence that was genuinely predicted, strongly discriminates alternatives, survives independent replication and holds across appropriate methods and domains.

26. What weakens a hypothesis most?

Repeated failure under severe, valid tests combined with competing hypotheses that predict the evidence better.

27. What is a hypothesis portfolio?

A set of live competing explanations maintained until evidence discriminates among them.

28. Why keep multiple hypotheses alive?

It prevents premature closure and encourages tests that distinguish rather than merely confirm.

29. Can AI generate hypotheses?

Yes. AI can propose patterns and candidate explanations, but scientific value still depends on domain plausibility, testability, evidence and independent evaluation.

30. Can AI test hypotheses?

AI can assist analysis, simulation and experiment design. But real-world hypotheses often require observations and interventions outside the model, plus safeguards against leakage and circular validation.

31. What is the final test of a hypothesis?

There is rarely one final test. Science accumulates evidence, probes boundaries, compares alternatives and revises models over time.

---

192. Where this page sits in the eduKateYishun scientific-method system

• How Scientific Measurement Works — how signals become measurements.
• How Scientific Models Work — how simplified representations organise mechanisms and predictions.
• How Scientific Evidence Works — how observations support claims.
• How Scientific Experiments Work — how discriminating tests are built.
• How Scientific Explanations Work — how mechanisms make sense of patterns.
• How Scientific Uncertainty Works — how remaining uncertainty is represented.
• How Scientific Prediction Works — how models commit to unseen outcomes.
• When Science Results Disagree — how conflicting results guide the next test.

193. Authoritative further reading

• University of California, Berkeley — Understanding Science: How Science Works
• National Academies / NCBI Bookshelf — Reproducibility and Replicability in Science
• National Academies / NCBI Bookshelf — Reference Manual on Scientific Evidence

194. How Scientific Hypotheses Work: the larger lesson

A hypothesis is where curiosity becomes accountable.

We notice something.

We propose what might explain it.

Then science asks us to do something psychologically difficult:

name the evidence that could show us we are wrong.

The hypothesis must generate expectations before the result, compete with alternatives, survive measurement and design, and change when evidence changes.

The strongest hypothesis culture is not one that celebrates being right.

It celebrates building ideas that can be tested honestly.

The complete loop is:

question → alternatives → hypothesis → prediction → discriminating test → evidence → update → next hypothesis

That loop is one of the engines of science.

---

195. Hypothesis Transformation Workbook

This workbook converts ordinary claims into stronger scientific hypotheses. Each example shows five layers: the vague idea, a testable version, the prediction it generates, one competing explanation and a result that would force revision. The purpose is not to memorise the wording. It is to learn how scientific claims acquire structure.

Transformation 1 — “Sunlight makes plants healthy”

Testable hypothesis: within the tested range, seedlings receiving a longer daily light period will show greater mean dry mass after 21 days than seedlings receiving a shorter light period under comparable water, soil and temperature conditions.

Prediction: the longer-light group will have higher mean dry mass.

Alternative: the lighting system also raises temperature.

Revision trigger: controlling temperature removes the difference.

Transformation 2 — “More water helps plants grow”

Testable hypothesis: increasing daily water supply from W1 to W2 improves measured growth only within a defined range.

Prediction: growth increases up to a point rather than indefinitely.

Alternative: excess water reduces root oxygen and harms growth.

Revision trigger: the relationship is non-linear or reverses above a threshold.

Transformation 3 — “Hotter water evaporates faster”

Testable hypothesis: with equal initial volume and exposed surface, water maintained at the higher tested temperature will lose more mass over the same interval.

Prediction: greater mass loss in the warmer condition.

Alternative: airflow differed.

Revision trigger: after controlling airflow, temperature no longer predicts mass loss.

Transformation 4 — “Exercise is good for the heart”

Testable hypothesis: a defined aerobic training programme changes a specified cardiovascular measure over a defined period relative to a comparison condition.

Prediction: the selected measure shifts in the expected direction.

Alternative: diet or medication changed simultaneously.

Revision trigger: the effect disappears in a controlled design.

Transformation 5 — “Sleep improves memory”

Testable hypothesis: participants receiving a defined sleep opportunity after learning will show higher delayed recall than participants kept awake for the same post-learning interval.

Prediction: higher delayed recall in the sleep condition.

Alternative: groups differ in fatigue during testing rather than memory consolidation.

Revision trigger: controlling test-time alertness eliminates the difference.

Transformation 6 — “Background music helps studying”

Testable hypothesis: under equal study time, instrumental music at a defined volume changes delayed comprehension performance relative to silence.

Prediction: specify increase, decrease or non-directional difference before testing.

Alternative: personal music preference moderates the effect.

Revision trigger: effect exists only for one preference group.

Transformation 7 — “Retrieval practice is better than rereading”

Testable hypothesis: with equal study time and content exposure, retrieval practice produces higher delayed transfer performance than rereading.

Prediction: larger difference on delayed transfer than on immediate familiarity.

Alternative: retrieval group receives more feedback.

Revision trigger: matched-feedback conditions remove the effect.

Transformation 8 — “Smaller classes are better”

Testable hypothesis: reducing instructional group size increases the frequency of individual feedback and student explanation, which in turn improves a defined learning outcome.

Prediction: feedback frequency mediates part of the outcome difference.

Alternative: tutor quality differs.

Revision trigger: group-size effect vanishes after tutor quality and feedback are controlled.

Transformation 9 — “Practice makes perfect”

Testable hypothesis: performance improves with additional correctly guided practice up to a plateau under a defined task and feedback regime.

Prediction: rapid early gains followed by diminishing returns.

Alternative: repeated practice raises familiarity without transfer.

Revision trigger: gains disappear on novel tasks.

Transformation 10 — “Stress causes poor grades”

Testable hypothesis: higher measured stress predicts lower subsequent assessment performance after accounting for prior attainment and study time.

Prediction: stress measured before the assessment relates to later performance.

Alternative: poor expected performance causes stress.

Revision trigger: temporal analysis supports the reverse direction.

Transformation 11 — “Pollution causes illness”

Testable hypothesis: higher long-term exposure to Pollutant P increases incidence of Outcome O in Population Q, conditional on measured confounders.

Prediction: exposure-response pattern.

Alternative: socioeconomic conditions explain both exposure and outcome.

Revision trigger: well-designed natural experiment shows no change despite large exposure reduction.

Transformation 12 — “Coffee keeps people awake”

Testable hypothesis: a defined caffeine dose administered at Time T changes a specified alertness or sleep measure relative to placebo.

Prediction: measurable difference in the expected direction.

Alternative: expectation produces the effect.

Revision trigger: blinded placebo-controlled trial shows no difference.

Transformation 13 — “The drug works”

Testable hypothesis: Drug D at Dose X improves Primary Outcome Y over Time T compared with control in Population P.

Prediction: a pre-specified difference or distribution shift.

Alternative: spontaneous recovery.

Revision trigger: controlled trial repeatedly finds no clinically meaningful benefit.

Transformation 14 — “This biomarker predicts disease”

Testable hypothesis: biomarker B measured before diagnosis improves out-of-sample risk prediction beyond baseline clinical variables.

Prediction: better calibrated predictive skill on external data.

Alternative: biomarker is measured after disease process is already obvious.

Revision trigger: performance collapses in prospective testing.

Transformation 15 — “The new treatment is just as good”

Testable hypothesis: the new treatment is non-inferior to the standard within a pre-specified clinically meaningful margin.

Prediction: confidence interval remains within the allowed margin.

Alternative: study lacks sensitivity to detect meaningful differences.

Revision trigger: assay sensitivity is inadequate or interval crosses the margin.

Transformation 16 — “The two sensors agree”

Testable hypothesis: measurement differences between Sensor A and B remain within a pre-specified tolerance across the validated range.

Prediction: bias and limits of agreement remain acceptable.

Alternative: agreement holds only at mid-range values.

Revision trigger: systematic divergence appears at extremes.

Transformation 17 — “The material is stronger”

Testable hypothesis: Material A has higher mean tensile strength than Material B under standardised specimen geometry and testing conditions.

Prediction: higher tensile-strength distribution.

Alternative: manufacturing batch differs.

Revision trigger: matched-batch tests eliminate the difference.

Transformation 18 — “The bridge vibrates because of wind”

Testable hypothesis: vibration amplitude increases when wind forcing approaches a structural resonance frequency.

Prediction: spectral peaks align with wind and structure frequencies.

Alternative: traffic excitation drives the vibration.

Revision trigger: vibration persists with traffic changes but not wind changes, or vice versa.

Transformation 19 — “The fan is causing overheating”

Testable hypothesis: reduced fan performance causes elevated operating temperature.

Prediction: restoring fan performance lowers temperature under the same load.

Alternative: thermal interface degradation.

Revision trigger: fan replacement does not change temperature while interface repair does.

Transformation 20 — “The software update caused the failures”

Testable hypothesis: failure rate increased after Version V because of a specific change in Module M.

Prediction: rollback or isolated module reversal reduces failures.

Alternative: traffic volume changed simultaneously.

Revision trigger: controlled rollback has no effect.

Transformation 21 — “The algorithm is biased”

Testable hypothesis: Model M shows systematically different error rates across Groups G1 and G2 under a defined fairness metric after accounting for relevant prevalence and measurement differences.

Prediction: disparity persists in held-out data.

Alternative: label quality differs across groups.

Revision trigger: correcting label quality removes the disparity.

Transformation 22 — “The model learned the disease”

Testable hypothesis: the model uses pre-diagnostic biological signal that generalises across sites.

Prediction: performance remains on external sites with different workflows.

Alternative: model learned site-specific process artifacts.

Revision trigger: external performance collapses despite similar disease prevalence.

Transformation 23 — “The trading strategy works”

Testable hypothesis: Strategy S produces positive risk-adjusted returns after realistic transaction costs on data unavailable during development.

Prediction: forward performance exceeds baseline.

Alternative: backtest overfit.

Revision trigger: repeated live or holdout tests fail.

Transformation 24 — “Customers prefer Product A”

Testable hypothesis: randomised exposure to Product A produces a higher selection probability than Product B in the target population.

Prediction: higher choice rate under random assignment.

Alternative: placement or price differs.

Revision trigger: controlled presentation removes the difference.

Transformation 25 — “The recommendation increased engagement”

Testable hypothesis: showing Recommendation R causally increases defined engagement relative to not showing it.

Prediction: randomised users receiving R show higher incremental engagement.

Alternative: R is shown to users already likely to engage.

Revision trigger: randomised effect is negligible.

Transformation 26 — “The species is moving because the climate is warming”

Testable hypothesis: distribution shifts track changes in temperature-defined habitat suitability more closely than alternative drivers.

Prediction: spatial shifts align with thermal niches across regions.

Alternative: land-use change.

Revision trigger: shifts occur where temperature is stable but habitat loss is severe.

Transformation 27 — “The conservation policy worked”

Testable hypothesis: Policy P increased population growth relative to comparable untreated areas or pre-policy counterfactual trends.

Prediction: improvement begins after intervention and is stronger where policy intensity is greater.

Alternative: favourable weather.

Revision trigger: matched comparison sites improve equally.

Transformation 28 — “The pathogen became more transmissible”

Testable hypothesis: Variant V has higher intrinsic transmission rate after accounting for immunity, behaviour and founder effects.

Prediction: consistent growth advantage across regions and contexts.

Alternative: immune escape alone.

Revision trigger: growth advantage disappears in comparable immune backgrounds.

Transformation 29 — “The fossil rewrites the timeline”

Testable hypothesis: the specimen genuinely belongs to Species S and derives from geological layer L of age A.

Prediction: independent dating and stratigraphic evidence agree.

Alternative: reworking or misidentification.

Revision trigger: independent methods contradict age or identity.

Transformation 30 — “The star has a planet”

Testable hypothesis: periodic brightness dips are caused by a transiting planet.

Prediction: stable periodicity, repeatable depth and supporting radial-velocity or spectral evidence where detectable.

Alternative: stellar activity.

Revision trigger: wavelength dependence and irregular timing favour starspots.

Transformation 31 — “Dark matter explains the rotation curve”

Testable hypothesis: an unseen mass distribution with specified profile accounts for observed dynamics and additional gravitational phenomena.

Prediction: lensing and other dynamical effects follow the same mass model.

Alternative: modified gravity model.

Revision trigger: new observations consistently favour a different model across independent phenomena.

Transformation 32 — “The earthquake was predictable”

Testable hypothesis: Precursor P predicts increased short-term probability of an earthquake beyond the regional baseline.

Prediction: prospectively registered P events are followed by earthquakes at a rate above baseline.

Alternative: P is common without earthquakes.

Revision trigger: low prospective skill and high false-alarm rate.

Transformation 33 — “This signal predicts an eruption”

Testable hypothesis: Signal S adds predictive skill for eruption within Window W beyond standard monitoring variables.

Prediction: calibrated probability improves out of sample.

Alternative: S only measures general unrest.

Revision trigger: no incremental predictive skill.

Transformation 34 — “The river cleared because of the regulation”

Testable hypothesis: Regulation R reduced pollutant concentration downstream relative to a counterfactual trend.

Prediction: regulated pollutants decline after enforcement, especially downstream of regulated sources.

Alternative: rainfall dilution.

Revision trigger: upstream and downstream sites change identically with rainfall.

Transformation 35 — “The lake lost oxygen because of algal blooms”

Testable hypothesis: nutrient-driven algal production followed by decomposition caused oxygen depletion.

Prediction: nutrient increase, algal biomass and later oxygen decline show the expected temporal sequence.

Alternative: thermal stratification alone.

Revision trigger: oxygen loss occurs without biomass or nutrient changes but tracks stratification.

Transformation 36 — “This treatment works only for high-risk patients”

Testable hypothesis: baseline risk moderates treatment effect.

Prediction: interaction between treatment and pre-specified risk measure.

Alternative: subgroup result is chance.

Revision trigger: prospective replication fails.

Transformation 37 — “The exam intervention improves understanding”

Testable hypothesis: intervention improves transfer to unseen problems, not merely immediate score on trained items.

Prediction: advantage persists on delayed novel tasks.

Alternative: intervention teaches answer patterns.

Revision trigger: gain disappears when surface format changes.

Transformation 38 — “The tutor fixed the misconception”

Testable hypothesis: the learner now applies the corrected model across delayed and changed contexts without prompting.

Prediction: no recurrence on fresh items.

Alternative: student memorised correction.

Revision trigger: misconception returns when representation changes.

Transformation 39 — “The student is careless”

Testable hypothesis: errors arise from a recurring process such as missing negative words, units or comparison targets.

Prediction: errors cluster around that process across topics.

Alternative: concept knowledge is unstable.

Revision trigger: process checklist does not reduce error recurrence.

Transformation 40 — “The child needs more practice”

Testable hypothesis: the current method is correct but not yet fluent, so additional varied retrieval will improve speed and accuracy.

Prediction: performance improves with practice without extra conceptual teaching.

Alternative: underlying concept is wrong.

Revision trigger: more practice repeats the same misconception.

---

196. Hypothesis Integrity Appendix

Keep the original hypothesis

When evidence arrives, preserve the wording of the original claim. Revised hypotheses should be versioned rather than silently replacing the old one. Otherwise hindsight can make the research programme look more predictive than it was.

Keep exploratory and confirmatory work visible

Exploration can generate important ideas. Confirmation can test them cleanly. Both are valuable. The integrity problem appears when the boundary is hidden.

Keep failed tests

Negative and contradictory results belong in the scientific record. Removing them makes weak hypotheses look stronger and encourages repeated wasted effort.

Keep the alternatives

When one hypothesis is preferred, document which alternatives were considered and why they lost support. This makes the reasoning auditable and helps future researchers revisit assumptions if new evidence appears.

Keep the domain

A hypothesis supported in one range, population or environment should not silently expand into a universal law. Generalisation is itself another hypothesis requiring evidence.

Keep the distinction between support and truth

Scientific evidence can strongly support a hypothesis. That does not convert empirical inference into mathematical proof. Future observations can expose limits, exceptions or better models.

Keep the failure criterion

A useful hypothesis should retain a clear route to losing support. If revisions make the hypothesis compatible with every possible outcome, scientific content has been drained away.

Keep the next test prospective

After a hypothesis is revised, the strongest evidence comes from a test designed and evaluated after the revision, not by showing that the new version fits the old data from which it was invented.

197. Scope of this guide

This guide focuses on candidate scientific claims and how they are exposed to failure. Related topics are treated separately so each can be examined in enough depth:

• models focus on representation;
• measurement focuses on quantified observation;
• evidence focuses on confidence;
• experiments focus on test construction;
• explanations focus on mechanism;
• uncertainty focuses on ranges and unknowns;
• prediction focuses on forward forecasts.

Hypothesis reasoning connects those systems by asking one recurring question:

What idea are we putting at risk, and what evidence could make us change it?

198. The scientific hypothesis contract

A serious scientific hypothesis makes an implicit contract:

I state a candidate claim clearly enough that consequences can be derived. I identify the conditions under which the claim applies. I allow competing explanations. I specify what evidence would count against me. I will not rewrite the claim after the outcome without admitting the revision. And I will let repeated, discriminating evidence change what I believe.

That contract is the difference between having an idea and testing one scientifically.

---

199. The Scientific Hypothesis Notebook

A hypothesis notebook keeps the scientific reasoning visible across time. Its purpose is not decorative record-keeping. It prevents hindsight from quietly rewriting what was believed, what was predicted and what evidence changed the claim.

Notebook field 1 — Question

Write the phenomenon or problem that needs explanation. A good question does not assume the answer. “Why are these plants different?” is more open than “How did the fertiliser make these plants taller?” when fertiliser has not yet been established as the cause.

Notebook field 2 — Observation that triggered the question

Separate the initial observation from interpretation. Record what was actually measured, seen or detected. This prevents the hypothesis from being smuggled into the evidence statement.

Notebook field 3 — Candidate hypothesis

Write the proposed relationship or mechanism in one clear version. Add a version number and date. If the hypothesis changes later, preserve the old version.

Notebook field 4 — Domain

Record the intended population, range, environment, time scale and relevant conditions. A hypothesis without a domain can expand silently until one local finding becomes an unjustified universal claim.

Notebook field 5 — Operational definitions

Define how important concepts will be measured. If “growth” means dry mass in one study and height in another, the hypothesis tests may not be equivalent. The notebook should make that choice visible.

Notebook field 6 — Competing hypotheses

List at least one plausible alternative where possible. Include measurement artifact or selection explanations when relevant. This makes the investigation comparative rather than purely confirmatory.

Notebook field 7 — Auxiliary assumptions

Record background assumptions needed to derive the prediction. Examples include instrument calibration, treatment delivery, sample comparability and model validity.

Notebook field 8 — Prediction

Write what should be observed if the hypothesis and assumptions are correct. The prediction should be recorded before the key outcome is examined.

Notebook field 9 — Discriminating observation

State the observation that would favour one hypothesis over another. If no such observation can be identified, the planned test may not be informative enough.

Notebook field 10 — Failure criterion

Write what result would reduce confidence in the hypothesis. A failure criterion does not need to mean instant rejection, but it should make clear that the idea can lose support.

Notebook field 11 — Study design

Record the comparison, controls, randomisation, blinding, sampling and other design features relevant to the hypothesis. Design choices should follow the alternative explanations that need to be managed.

Notebook field 12 — Analysis plan

For confirmatory work, state the main outcome, statistical model and decision rule before seeing the relevant results. Exploratory analyses can be added later and labelled as such.

Notebook field 13 — Result

Record the result without immediately rewriting the hypothesis. Keep description of evidence separate from interpretation.

Notebook field 14 — Evidence update

State whether the evidence increased, decreased or left unchanged the credibility of each candidate hypothesis. Explain why.

Notebook field 15 — Revision

If the hypothesis changes, write Version 2 beside Version 1. Explain what evidence motivated the change and what new prediction Version 2 makes.

Notebook field 16 — Replication status

Record whether the result has been independently replicated, conceptually replicated or only observed in the original study. One result and a replicated result should not be described with the same confidence.

Notebook field 17 — Boundary status

Record where the hypothesis has been tested successfully and where it has failed or remains untested. Boundaries are scientific knowledge, not embarrassments.

Notebook field 18 — Alternative-hypothesis status

Mark which alternatives were weakened, which remain viable and whether new alternatives emerged. A preferred hypothesis is stronger when serious alternatives have faced discriminating evidence.

Notebook field 19 — Next test

Identify the most informative next observation or experiment. The next test should ideally target the biggest remaining uncertainty rather than merely repeating the easiest successful condition.

Notebook field 20 — Retirement rule

State what pattern of evidence would justify abandoning or substantially replacing the hypothesis. Retirement criteria help prevent endless ad hoc rescue.

---

200. Twenty audit questions before calling a hypothesis “supported”

1. Was the hypothesis stated before the key result was known?
2. Were important concepts operationalised clearly?
3. Was the population or domain specified?
4. Were plausible alternative hypotheses considered?
5. Did the hypothesis generate a prediction that could have failed?
6. Did competing hypotheses make different predictions?
7. Was the test capable of detecting the predicted effect?
8. Were relevant controls or comparison groups appropriate?
9. Was measurement reliable and valid enough for the claim?
10. Were analysis choices planned or transparently labelled exploratory?
11. Was the result large enough to matter scientifically, not merely statistically?
12. How uncertain is the estimated effect?
13. Could confounding, selection or reverse causation explain the pattern?
14. Was the result tested on new data or only the discovery dataset?
15. Did independent replication succeed?
16. Do different methods converge on the same conclusion?
17. Has the hypothesis survived tests near its proposed boundaries?
18. What important alternatives remain?
19. What evidence would still make us reduce confidence?
20. What is the next severe test?

If several of these questions cannot be answered, the correct scientific language may be “preliminary”, “suggestive” or “requires further testing” rather than “established”.

201. Twenty audit questions after a hypothesis fails

1. Was the prediction derived correctly from the hypothesis?
2. Were the auxiliary assumptions reasonable?
3. Did the instrument work correctly?
4. Was the treatment or intervention implemented correctly?
5. Was the sample appropriate?
6. Was the measurement sensitive enough?
7. Was the study powered for the predicted effect?
8. Did a confounder differ between conditions?
9. Did the system enter a different regime?
10. Was the domain narrower than assumed?
11. Did the effect depend on an unmodelled moderator?
12. Did data processing introduce error?
13. Was the outcome definition changed?
14. Was the statistical model appropriate?
15. Did the competing hypothesis predict the result better?
16. Does independent evidence point to the same failure?
17. Would repetition under corrected conditions still test the original claim?
18. Does the hypothesis need narrowing rather than total rejection?
19. Does a new mechanistic hypothesis explain both the old successes and new failure?
20. What prospective test can distinguish the revised possibilities?

The purpose of this audit is not to save the favourite idea. It is to diagnose what the evidence actually challenged.

---

202. Hypothesis traffic-light system

Green — strong provisional support

• clear pre-specified hypothesis;
• discriminating predictions;
• strong measurement;
• appropriate design;
• meaningful effect;
• successful independent replication;
• alternative explanations weakened;
• known domain and boundaries.

Green does not mean “proven forever”. It means the hypothesis currently deserves substantial confidence within its tested domain.

Amber — plausible but incomplete

• some supportive evidence;
• important alternatives remain;
• effect imprecise;
• replication limited;
• domain unclear;
• test not highly discriminating.

Amber calls for better measurement, replication or a more informative test.

Red — weak or actively contradicted

• repeated severe failures;
• better competing explanations;
• support depends on ad hoc rescue;
• result fails replication;
• claim relies on biased or invalid measurement;
• hypothesis cannot state a possible failure.

Red means the idea should be revised, restricted or retired unless genuinely new evidence changes the picture.

---

203. From classroom hypothesis to scientific research programme

A classroom hypothesis often begins with one variable and one outcome.

A research programme can grow that idea through stages:

1. Existence: Is there a reproducible relationship?
2. Magnitude: How large is it?
3. Boundary: Where does it hold?
4. Mechanism: How does it occur?
5. Moderator: When does it become stronger or weaker?
6. Causality: Would intervention change the outcome?
7. Prediction: Does it forecast new cases?
8. Generality: Does it transfer across populations and methods?
9. Integration: How does it fit into a broader theory?

This progression explains why mature science rarely ends with “the hypothesis was correct”. Each answer creates more precise hypotheses.

204. Hypothesis progression example: temperature and reaction rate

Stage 1: Does reaction rate differ across temperature?

Stage 2: What is the quantitative relationship within the tested range?

Stage 3: Where does the relationship stop being monotonic?

Stage 4: Which molecular mechanism explains the change?

Stage 5: How does catalyst type moderate the temperature relationship?

Stage 6: Can the mechanistic model predict rate under a new temperature and concentration combination?

The same research question evolves from descriptive comparison into a predictive mechanistic theory.

205. Hypothesis progression example: learning method

Stage 1: Does retrieval practice improve delayed performance?

Stage 2: Does the advantage persist when total study time is equal?

Stage 3: Does it transfer to novel questions?

Stage 4: Which feedback conditions matter?

Stage 5: Does prior knowledge moderate the effect?

Stage 6: Which cognitive mechanism explains the benefit?

Stage 7: Can the model predict which students benefit most?

This is how a useful hypothesis becomes a programme rather than a one-off result.

206. Hypothesis progression example: conservation intervention

Stage 1: Did population growth change after protection?

Stage 2: Was the change larger than in comparable unprotected sites?

Stage 3: Which mechanism—reduced mortality, increased breeding or migration—explains the effect?

Stage 4: Which habitat conditions moderate effectiveness?

Stage 5: Does the intervention work in other regions?

Stage 6: Can the mechanistic model forecast long-term population response?

---

207. Hypothesis craftsmanship: writing claims that deserve tests

Good hypothesis writing is not about sounding formal. It is about exposing structure.

A well-crafted hypothesis often contains:

• entities or variables;
• the proposed relationship;
• direction when justified;
• domain or population;
• time or scale where relevant;
• a measurable implication.

But not every hypothesis needs every element in one sentence. Complex claims can be expressed through equations, causal diagrams or model specifications.

208. Avoid universal verbs without evidence

Words such as:

• always;
• never;
• all;
• only;
• must;

create strong universal claims. Use them only when the hypothesis truly intends that breadth and the science justifies it.

209. Avoid vague verbs

Words such as:

• helps;
• affects;
• improves;
• influences;
• works;

can hide the actual relationship.

Ask:

• increases what?
• decreases what?
• changes which distribution?
• over what time?
• compared with what?

210. Avoid mechanism claims without mechanism evidence

If the design only tests whether X and Y differ, do not write a hypothesis claiming that X acts through Mechanism M unless M is also measured or manipulated.

The claim should match the evidence the study can produce.

211. Avoid using the statistical test as the scientific hypothesis

“There will be a significant difference” is usually a poor scientific hypothesis.

It says nothing about:

• mechanism;
• magnitude;
• direction;
• scientific importance.

Write the substantive scientific claim first. Choose the statistical test afterward.

212. Avoid circular hypotheses

Weak:

“Students who are better learners will learn better.”

The predictor and outcome partly define one another.

Scientific hypotheses should avoid circular definitions that make confirmation automatic.

213. Avoid hypothesis–measurement identity

If “intelligence” is defined as the score on Test T, then a hypothesis that “intelligence predicts Test T” is partly circular.

Construct validity matters.

214. Avoid using absence of evidence as evidence of impossibility

If a study fails to detect Effect E, ask whether it was capable of detecting E.

Strong negative hypotheses require sensitive tests.

215. Avoid making every anomaly a new exception

Anomaly-driven revision is useful when it creates a broader, independently testable rule.

If every anomaly produces its own special clause, the hypothesis may be losing explanatory coherence.

---

216. Final hypothesis checklist for students

• I can state the question without assuming the answer.
• I can name at least one plausible alternative.
• I can define the important variables.
• I can say where the hypothesis should apply.
• I can derive a prediction before seeing the outcome.
• I can say what observation would weaken the hypothesis.
• I understand why the comparison is fair.
• I can separate observation from explanation.
• I will not call a hypothesis “proved” from one result.
• I can revise the claim if evidence disagrees.

217. Final hypothesis checklist for researchers

• Is the hypothesis substantive rather than merely statistical?
• Was it pre-specified or transparently labelled exploratory?
• Are operational definitions valid?
• Are major alternative hypotheses represented?
• Does the test discriminate among them?
• Are auxiliary assumptions documented?
• Is the design capable of detecting scientifically meaningful effects?
• Are multiplicity and researcher degrees of freedom controlled or disclosed?
• Are effect sizes and uncertainty reported?
• Is the analysis reproducible?
• Can the result be independently replicated?
• Are boundaries and moderators investigated?
• Does revision generate a new prospective prediction?
• Is there a rational retirement criterion?

218. The final standard

A scientific hypothesis earns its place not because it is clever, elegant, popular or written in technical language.

It earns its place because it creates a disciplined relationship between an idea and possible evidence.

The idea must risk something.

The evidence must be allowed to matter.

The alternatives must be taken seriously.

And the claim must remain capable of change.

A hypothesis is science when it is clear enough to test, vulnerable enough to fail, and useful enough to generate the next question.

---

219. Hypothesis Versioning and Generalisation Standard

A hypothesis should have a history. Version 1 is the claim before the first decisive test. Version 2 may narrow the domain after a boundary is discovered. Version 3 may add a moderator or replace the proposed mechanism. Keeping those versions visible prevents hindsight from making scientific progress look cleaner than it really was.

Versioning also separates revision from goalpost moving. A legitimate revision says what changed, why it changed, and what new evidence could test the revised claim. Goalpost moving quietly changes the meaning of success after the outcome so that the idea can never lose.

Generalisation should be treated as a new hypothesis. If a relationship works in one school, one species, one laboratory, one material range or one population, the claim that it also works elsewhere requires evidence. Replication across contexts is therefore not repetitive bureaucracy. It tests the scope of the idea.

A useful generalisation sequence is:

1. establish the relationship in the original domain;
2. replicate under similar conditions;
3. change one important context;
4. measure whether the effect changes;
5. identify moderators and boundaries;
6. update the hypothesis domain;
7. test the broader claim prospectively.

This sequence protects science from two opposite errors. The first is overgeneralisation: assuming one successful result applies everywhere. The second is underlearning: treating every context difference as unrelated and never building a broader principle.

The strongest hypotheses become more general only when they survive deliberate attempts to transport them across methods, populations, ranges and environments.

220. The final hypothesis-owner rule

Future expansion of this page should preserve one governing boundary: the hypothesis owner is about how candidate scientific claims are formed, exposed to alternatives, made testable, revised, replicated and, when necessary, retired.

Its durable standard is:

State the idea clearly. Derive consequences before the answer. Build tests that alternatives can lose. Preserve failures. Revise transparently. Generalise only after the new domain has earned it.

That is how a scientific hypothesis survives not by being protected from failure, but by repeatedly earning another test.

Integrity close: A hypothesis deserves confidence only while evidence remains capable of changing it. The scientific habit is therefore not defending an idea at all costs, but designing fair opportunities for reality to disagree, recording those disagreements honestly, and building the next version from what the test actually taught us.