How Scientific Causation Works | From Correlation and Mechanism to a Claim About What Changes What

Scientific causation asks whether changing one thing would change another. Two variables can move together without one causing the other. This guide explores how scientists use comparisons, experiments, mechanisms and counterfactual reasoning to assess cause and effect while checking alternative explanations.

A causal claim is stronger than an association. “X and Y are related” says they vary together. “X causes Y” says something counterfactual: if we intervened on X while holding the relevant causal structure appropriately fixed, Y would change because of that intervention. The difference sounds small in ordinary language, but scientifically it is enormous. Correlation can arise from direct causation, reverse causation, confounding, selection, measurement, shared causes, feedback or chance.

Work through the examples to define a causal question, identify possible confounding and assess what a comparison can support. Related guides explain evidence, experiments, mechanisms and explanations, hypotheses, prediction and theories.

---

Choose your route

• Basic route: correlation, cause, reverse causation and confounding.
• School Science route: fair tests, changed variables, controlled conditions and cause-and-effect explanations.
• Experiment route: randomisation, controls, blinding and why intervention strengthens causal inference.
• Observational route: confounding, selection bias, natural experiments and longitudinal evidence.
• Mechanism route: mediation, pathways, necessary causes, sufficient causes and causal chains.
• Modern causal-inference route: counterfactuals, potential outcomes, causal diagrams, backdoor paths and target estimands.
• Decision route: when a predictive association is enough and when an intervention requires causal knowledge.

1. Causation asks a stronger question than association

Suppose two variables move together.

Examples:

• people carrying umbrellas are more likely to be near rain;
• children with larger shoe sizes have larger vocabularies;
• students who attend more tutoring may score higher;
• cities with more firefighters may have more fire damage;
• patients taking a drug may have worse health outcomes.

Each relationship is an association.

None, by itself, tells us what causes what.

2. The umbrella example

Umbrella use and rain are associated.

But umbrellas do not cause rain.

Rain causes umbrella use.

This is reverse causation.

3. The shoe-size example

Among schoolchildren, larger shoe size can correlate with larger vocabulary.

Neither causes the other.

Age influences both.

This is confounding by a common cause.

4. The firefighter example

Larger fires attract more firefighters and also produce more damage.

Firefighter count and damage can therefore be positively associated even if firefighters reduce damage relative to what would have occurred without them.

The common cause is fire severity.

5. The drug example

Patients taking a treatment may have worse outcomes because the sickest patients are more likely to receive it.

This is confounding by indication.

The observed association can point in the opposite direction from the treatment’s causal effect.

6. Correlation can arise from several structures

If X and Y are associated, possibilities include:

1. X causes Y;
2. Y causes X;
3. Z causes both X and Y;
4. X and Y influence each other;
5. selection creates the association;
6. measurement creates the association;
7. chance creates the association;
8. several of these act together.

Causal inference is the discipline of distinguishing these structures.

7. Statistical significance does not establish cause

A small p-value can tell us that an observed association would be unusual under a specified null model.

It does not tell us:

• which direction causation runs;
• whether a confounder exists;
• whether selection bias created the pattern;
• whether the measurement is valid;
• whether the effect is large enough to matter.

The CDC’s Field Epidemiology Manual explicitly notes that statistical significance alone does not establish cause-and-effect. Chance, selection bias, information bias, confounding and other error must be considered before causal interpretation.

8. Temporal order matters

A cause must occur before its effect.

This sounds obvious.

But many datasets measure variables at the same time.

If sleep problems and anxiety are measured on one survey date, an association does not identify whether:

• sleep affects anxiety;
• anxiety affects sleep;
• both directions operate;
• another factor affects both.

Longitudinal data can help clarify order, though confounding can remain.

9. Causation is fundamentally counterfactual

To say X caused Y often means:

Y would have been different if X had been different, under otherwise relevantly comparable conditions.

This is a counterfactual claim.

For one individual, we usually cannot observe both worlds:

• what happened after treatment;
• what would have happened to the same individual at the same time without treatment.

This missing counterfactual is the central problem of causal inference.

10. The fundamental causal problem

Imagine a patient receives Drug A and recovers.

Did the drug cause recovery?

We cannot rewind the same patient to the same moment and observe the outcome without Drug A.

Science therefore constructs comparison groups intended to approximate the unobserved counterfactual.

11. Randomisation builds a comparison

In a randomized experiment, treatment assignment is determined by chance.

When performed well and with adequate sample size, randomisation helps make treatment groups comparable on both measured and unmeasured pre-treatment factors.

The intended systematic difference becomes the assigned intervention.

This is why randomized trials are such powerful tools for causal inference.

12. Randomisation does not make every problem disappear

Even randomized studies can suffer from:

• non-adherence;
• dropout;
• measurement bias;
• treatment crossover;
• missing data;
• chance imbalance in small samples;
• poor generalisability.

Randomisation strengthens one crucial link. It does not guarantee perfect study execution.

13. Causal inference needs a target

“What is the causal effect?” is incomplete.

Ask:

• effect of what intervention?
• compared with what?
• for whom?
• over what time?
• on what outcome?

This full specification is sometimes called the causal estimand.

14. Treatment versus treatment

A drug’s effect can differ depending on comparison:

• drug vs placebo;
• drug vs no treatment;
• drug vs standard treatment;
• high dose vs low dose.

The causal question must specify the contrast.

15. Population matters

A treatment effect in:

• children;
• older adults;
• severely ill patients;
• healthy volunteers;

may differ.

Causal effects are properties of interventions in populations under conditions, not universal magic numbers.

16. Time matters

An intervention can:

• help in the short term;
• harm in the long term;
• show delayed benefit;
• show early side effects.

Always define the follow-up horizon.

---

17. Necessary causes and sufficient causes

A necessary cause must be present for the outcome to occur.

A sufficient cause is a set of conditions that can produce the outcome.

Many real-world outcomes are multicausal.

No single factor may be necessary or sufficient by itself.

18. Multi-causal systems

Disease can result from combinations of:

• genetics;
• exposure;
• behaviour;
• age;
• environment;
• chance biological events.

Asking “What is the cause?” may therefore be less useful than asking:

“Which causal factors changed risk, and under what conditions?”

19. Causal chains

Causes often act through sequences.

Example:

air pollution → airway inflammation → physiological damage → symptoms

Each arrow is a causal claim.

The full chain becomes stronger when evidence supports the intermediate links.

20. Mediation

A mediator lies on the path from cause to outcome.

If X affects M and M affects Y, M may mediate part of X’s effect on Y.

Mediation helps answer:

“How does the effect happen?”

It belongs at the intersection of causal inference and mechanism.

21. Mediators are not confounders

A confounder is a pre-treatment common cause of exposure and outcome.

A mediator occurs after the exposure on the causal pathway.

Adjusting for a mediator can remove part of the effect we are trying to estimate.

This distinction is crucial.

22. Moderation or effect modification

A moderator changes the strength or direction of a causal effect.

Examples:

• a drug works differently by genotype;
• a teaching method works differently by prior knowledge;
• heat stress affects materials differently by composition.

The causal effect is therefore heterogeneous.

23. Interaction

Two causes can interact.

The effect of X can depend on Z.

This means causal systems are not always additive.

Interaction can be:

• biological;
• statistical;
• mechanistic;
• decision-specific.

24. Mechanism strengthens causal understanding

An association becomes more convincing when scientists can trace:

• how exposure changes a process;
• how the process changes the outcome;
• how blocking the process alters the effect.

Mechanism is not always required for causal inference, but it can greatly strengthen and deepen it.

25. Mechanism alone is not enough

A plausible mechanism can still be wrong.

History contains many biologically plausible claims that failed empirical testing.

Mechanism must meet evidence.

26. Randomised experiments and mechanism complement each other

Randomisation strengthens the causal contrast.

Mechanistic evidence helps explain:

• why the intervention works;
• which subgroups benefit;
• where the effect should fail;
• how to generalise beyond the original experiment.

---

27. Confounding: the classic alternative explanation

A confounder is a factor that helps create a misleading association between exposure and outcome because it influences or is associated with both in the relevant causal structure.

Classic teaching example:

• coffee drinking associated with disease;
• smoking associated with coffee drinking;
• smoking causes disease.

If smoking is not controlled, coffee can look causally harmful even if the relationship is partly or entirely confounded.

28. Confounding is a causal concept, not just a statistical one

A variable is not a confounder merely because including it changes a regression coefficient.

Confounding depends on causal structure.

Ask:

• does the variable precede the exposure?
• does it affect the outcome?
• does it influence exposure?
• is it outside the causal pathway?

29. Overadjustment

Adjusting for every available variable can make causal inference worse.

You might adjust for:

• a mediator;
• a collider;
• a variable affected by treatment.

These adjustments can create bias.

30. Colliders

A collider is a variable caused by two other variables.

Conditioning on a collider can create an association between its causes even when none existed before.

Example structure:

X → C ← Y

If we select or adjust based on C, X and Y can become statistically associated.

31. Selection bias can behave like collider bias

Suppose hospital admission depends on both Exposure X and Disease Y.

If we study only admitted patients, conditioning on admission can create or distort an association between X and Y.

Selection is part of causal structure.

32. Causal diagrams

Directed acyclic graphs, often called DAGs, represent assumed causal relationships using arrows.

They help researchers reason about:

• confounders;
• mediators;
• colliders;
• selection;
• which variables to adjust for.

The graph does not prove the causal structure.

It makes assumptions explicit enough to inspect.

33. A DAG is a hypothesis about causation

The arrows represent causal assumptions.

Different scientists can draw different graphs.

Evidence and domain knowledge are needed to choose among them.

DAGs are reasoning tools, not automatic truth machines.

34. Backdoor paths

In causal diagrams, a backdoor path is a non-causal route that can create association between exposure and outcome.

Confounder adjustment aims to block these paths appropriately.

35. Do not block the causal path accidentally

If the question is the total effect of X on Y, adjusting for a mediator can remove part of the very effect being estimated.

Causal adjustment follows the estimand.

36. No universal “control for everything” rule

Good causal analysis asks:

Which variables should be controlled for given the causal question and assumed graph?

That is a theory-guided question.

37. Longitudinal evidence

Following exposure and outcome through time can clarify:

• temporal order;
• dose and duration;
• lag effects;
• changes within individuals.

But longitudinal studies can still suffer from confounding, dropout and time-varying bias.

38. Time-varying confounding

Some confounders change over time and are themselves affected by earlier treatment.

Ordinary regression adjustment can then produce bias.

Modern causal methods such as marginal structural models were developed for such settings.

39. Natural experiments

Sometimes nature, policy or administrative rules create exposure differences that approximate random assignment.

Examples can include:

• policy thresholds;
• lotteries;
• geographic boundaries;
• timing changes;
• unexpected environmental shocks.

Natural experiments can strengthen causal inference when their assumptions are credible.

40. Natural does not mean automatically valid

The causal leverage comes from the assignment mechanism.

Ask:

• why were some units exposed and others not?
• could units manipulate assignment?
• did other factors change at the same time?
• is the comparison local or general?

41. Regression discontinuity

When treatment changes at a threshold, units just above and below the cut-off may be comparable.

Example:

a programme begins for students whose score falls below a threshold.

If assignment cannot be manipulated precisely, the threshold can create a quasi-experiment.

42. Difference-in-differences

Difference-in-differences compares:

• change over time in an exposed group;
• with change over time in a comparison group.

A key assumption is that, without treatment, the groups would have followed parallel trends.

That assumption must be argued and tested as far as possible.

43. Instrumental variables

An instrumental variable influences exposure but affects the outcome only through that exposure under specific assumptions.

Valid instruments can estimate causal effects when confounding is otherwise difficult.

But the assumptions are strong and often untestable directly.

44. Matching

Matching attempts to make exposed and unexposed groups more comparable on measured covariates.

It can reduce measured confounding.

It cannot guarantee control of unmeasured confounding.

45. Stratification

Stratification compares exposure-outcome relationships within levels of a confounder.

This can reduce confounding if the relevant factor is measured well.

46. Regression adjustment

Regression can adjust for measured covariates.

But statistical adjustment does not automatically make an observational study causal.

The causal assumptions come first.

47. Propensity scores

A propensity score summarises measured covariates into the estimated probability of exposure or treatment.

It can support:

• matching;
• weighting;
• stratification;
• adjustment.

It does not solve unmeasured confounding by itself.

48. Inverse probability weighting

Weighting can create a pseudo-population in which exposure is less associated with measured confounders.

When assumptions hold, causal effects can be estimated.

Poorly estimated or extreme weights can make results unstable.

49. Target trial thinking

One modern causal strategy asks researchers to imagine the ideal randomized trial they would run if it were feasible.

Then observational data are analysed to emulate that trial as closely as possible.

This clarifies:

• eligibility;
• treatment strategies;
• time zero;
• follow-up;
• outcome;
• causal contrast.

50. Time zero matters

Exposure assignment, eligibility and follow-up should align.

Misalignment can create immortal-time bias and other distortions.

Causal analysis is often about getting the timeline right.

---

51. Potential outcomes: a formal counterfactual language

Modern causal inference often represents causal effects using potential outcomes.

For one unit:

• Y(1) = outcome if treated;
• Y(0) = outcome if untreated.

The individual causal effect would be:

Y(1) − Y(0)

But only one of these outcomes can usually be observed for the same unit at the same time.

This formalises the fundamental problem of causal inference.

52. Average treatment effect

Because individual counterfactuals are unobservable, science often targets average causal effects.

The average treatment effect asks:

On average in the target population, how would the outcome differ if everyone received treatment versus if everyone received control?

This is a population-level counterfactual.

53. Average treatment effect on the treated

Sometimes we ask:

“What was the average effect among those who actually received treatment?”

This can differ from the population average if treatment effects are heterogeneous or treatment selection is non-random.

54. Intention-to-treat effect

In randomized trials, the intention-to-treat effect compares groups according to assigned treatment, regardless of adherence.

It answers a policy-relevant question:

“What is the effect of assigning this intervention under real trial conditions?”

This can differ from the effect of actually receiving treatment.

55. Per-protocol effect

A per-protocol effect asks about following the treatment strategy as specified.

Because adherence is not randomized, estimating this effect can require additional causal assumptions.

The estimand must match the question.

56. No single “treatment effect” exists without definition

Possible effects include:

• assignment effect;
• treatment received effect;
• dose effect;
• effect among treated;
• effect among eligible population;
• short-term effect;
• long-term effect.

Causal language should be precise enough that two researchers know they are estimating the same quantity.

57. Consistency

Potential-outcome reasoning often assumes consistency:

If a person actually receives Treatment A in the defined way, their observed outcome corresponds to the potential outcome under A.

This sounds simple, but treatment definitions can be vague.

“Exercise” can mean many programmes.

“Tutoring” can mean many instructional systems.

Causal effects require well-defined interventions.

58. Positivity

Positivity means that, for the covariate patterns relevant to the target population, there is a nonzero chance of receiving each treatment being compared.

If one group always receives treatment and another never can, data cannot directly compare counterfactual strategies in those strata.

59. Exchangeability

Exchangeability means the treatment groups are comparable with respect to potential outcomes, either because of randomisation or because confounders have been controlled adequately under assumptions.

Randomisation creates exchangeability by design in expectation.

Observational studies must argue for it using measured variables and causal knowledge.

60. No unmeasured confounding

A common observational assumption is:

after conditioning on the chosen covariates, there are no remaining common causes of treatment and outcome.

This assumption is powerful and cannot usually be proven from the observed data alone.

Domain knowledge matters.

61. Identification

A causal effect is identified when, under the stated causal assumptions, it can be expressed using the observed data distribution.

Identification is distinct from estimation.

First ask:

“Can the causal effect be learned from this study design under our assumptions?”

Then ask:

“How precisely can we estimate it?”

62. Estimation

Once identified, causal effects can be estimated using statistical methods.

Methods may include:

• regression;
• matching;
• weighting;
• g-computation;
• targeted maximum likelihood;
• instrumental variables.

Different estimators can target the same causal estimand.

63. Identification assumptions are not solved by complex algorithms

A machine-learning estimator can reduce modelling bias or improve flexibility.

It cannot recover a causal effect if the required causal variables were never observed and the identifying assumptions are false.

Better statistics do not remove the need for causal design.

---

64. Randomized controlled experiments

Randomisation is powerful because treatment assignment is separated from pre-treatment causes of the outcome.

In expectation, randomisation balances:

• known confounders;
• unknown confounders;
• measured factors;
• unmeasured factors.

This gives causal interpretation a strong foundation.

65. Randomization creates groups comparable before treatment

Imagine assigning a large number of participants to treatment or control by a fair random mechanism.

Before treatment, neither group was selected because of prognosis.

Any systematic outcome difference after treatment is therefore less plausibly explained by baseline confounding.

66. Randomisation is a design property

It is not enough that groups “look balanced” afterward.

Randomisation creates a known assignment mechanism.

That assignment mechanism supports valid probability statements and causal contrasts.

67. Small randomised trials can still be imbalanced

Randomisation balances groups on average across repeated assignments.

In one small trial, important covariates can differ by chance.

This affects precision and sometimes interpretation.

68. Blocking and stratified randomisation

Researchers can randomise within important strata to improve balance.

Examples:

• age group;
• school;
• baseline severity;
• site.

Design can preserve randomisation while improving comparability.

69. Cluster randomisation

Sometimes whole groups are randomized:

• schools;
• villages;
• hospitals;
• clinics.

Outcomes within clusters are correlated.

Analysis must respect the cluster assignment.

70. Randomisation can answer policy questions directly

If a ministry is deciding whether to assign a programme, the intention-to-treat effect can be exactly the relevant causal quantity.

Not every question requires a hypothetical perfectly adherent intervention.

71. Blinding reduces post-randomisation bias

Blinding can reduce:

• expectation effects;
• differential measurement;
• differential treatment;
• selective outcome assessment.

Randomisation handles baseline confounding.

Blinding addresses other pathways.

72. Placebo controls

Placebos can help separate:

• specific treatment effect;
• expectation;
• attention;
• context.

The appropriate control depends on the causal question.

73. Active controls

When withholding treatment is unethical, the comparison can be:

• new treatment;
• standard treatment.

The causal question becomes comparative effectiveness.

74. Randomisation does not ensure generalisability

A trial can estimate an internal causal effect very well for its participants yet generalise poorly to another population.

External validity is a separate problem.

75. Trial participants may differ

Eligibility criteria can exclude:

• older patients;
• people with multiple conditions;
• people unable to attend follow-up;
• different geographic groups.

Transporting causal effects requires additional assumptions.

---

76. Observational causal inference

Many important causal questions cannot be randomized.

Examples include:

• smoking and lung cancer;
• air pollution;
• long-term climate exposure;
• education policy already implemented;
• rare harms;
• natural disasters.

Observational studies therefore remain essential.

77. Observational does not mean non-causal

An observational design can support strong causal inference when:

• temporal order is clear;
• confounders are measured and controlled;
• bias is limited;
• evidence replicates across designs;
• mechanism supports the relationship;
• natural experiments provide leverage.

The burden is heavier because treatment assignment was not controlled by researchers.

78. Smoking and lung cancer

Randomizing people to smoke for decades would be unethical.

Yet multiple lines of observational, mechanistic, dose-response and cessation evidence established smoking as a cause of lung cancer.

This is a classic case showing that randomisation is powerful but not the only route to causal knowledge.

79. Converging observational designs

Confidence rises when:

• cohort studies;
• case-control studies;
• natural experiments;
• mechanistic studies;
• different populations;

all point toward the same causal interpretation.

80. Triangulation

Triangulation deliberately combines methods with different biases.

If each method has different weak points yet reaches a similar conclusion, one shared error is less likely to explain everything.

81. Negative controls

A negative control is an exposure or outcome that should not be causally affected in the hypothesised way.

If an apparent effect also appears in the negative control, hidden bias may be present.

82. Positive controls

A positive control checks whether the study can detect a relationship known to exist.

This can reveal whether the design has enough sensitivity.

83. Sensitivity analysis

Observational causal inference should ask:

“How strong would an unmeasured confounder need to be to explain away the result?”

Sensitivity analyses do not prove no confounding exists.

They quantify how vulnerable the conclusion is.

84. E-values and other robustness metrics

Some methods summarise the minimum strength of unmeasured confounding needed to explain an association.

These metrics can aid interpretation but do not replace subject-matter judgment.

85. Instrumental variables revisit the assignment problem

A credible instrument behaves like a source of treatment variation unrelated to other causes of outcome, except through treatment.

The challenge is defending the assumptions.

Weak or invalid instruments can produce misleading estimates.

86. Mendelian randomisation

In genetic epidemiology, inherited genetic variants can sometimes serve as instruments for exposures.

The method uses random allocation of alleles at conception as a source of quasi-random variation.

But assumptions such as no horizontal pleiotropy must be considered.

87. Policy thresholds as quasi-random variation

If a rule assigns treatment based on a cut-off, individuals near the threshold can be nearly comparable.

This can support local causal estimates.

88. Interrupted time series

An interrupted time-series design examines whether the level or trend changes after an intervention beyond the pre-existing trajectory.

It can be useful for population-level policies where no untreated control exists.

Other simultaneous changes remain a threat.

89. Synthetic controls

Synthetic control methods combine untreated comparison units to construct a counterfactual trajectory for one treated unit.

They are useful for policy interventions affecting one region or country.

Validity depends on whether pre-treatment patterns and predictors create a credible counterfactual.

90. Fixed effects

Panel-data models can control for stable unmeasured differences between units by comparing within-unit changes.

They do not solve confounding from time-varying factors automatically.

91. Causal inference is design first, statistics second

The best statistical method cannot rescue a fundamentally incomparable design.

Before choosing regression, matching or weighting, specify:

• causal question;
• time zero;
• treatment strategies;
• outcome;
• confounders;
• selection process.

Then choose the estimator.

---

92. Bradford Hill considerations

In epidemiology, the Bradford Hill considerations are often used as guides when assessing whether an observed association may be causal.

They include ideas such as:

• temporality;
• strength of association;
• consistency;
• dose-response;
• plausibility;
• coherence;
• experiment or cessation evidence;
• specificity in some contexts.

They are not a mechanical checklist.

93. Temporality is essential

If X is proposed to cause Y, X must occur before Y.

This is the one criterion that is logically required for ordinary causal interpretation.

94. Strength of association

A large association can be harder to explain entirely by modest confounding.

But small causal effects can be real.

Strong association is not necessary.

95. Consistency

If different studies, populations and methods find similar relationships, confidence can increase.

But true effects can vary across contexts.

Perfect consistency is not required.

96. Dose-response

If more exposure produces more outcome, that can support causality.

But many causal relationships are:

• thresholded;
• saturating;
• U-shaped;
• non-monotonic.

Absence of simple dose-response does not rule out cause.

97. Biological plausibility

A mechanism consistent with current knowledge strengthens causal interpretation.

But scientific knowledge changes.

Historically, some true relationships seemed implausible before the mechanism was known.

98. Coherence

A proposed cause should fit reasonably with related biological or physical evidence.

Conflict can signal:

• bad causal claim;
• incomplete existing theory;
• new science.

99. Experiment and cessation

If reducing exposure reduces outcome, that can support causality.

Examples:

• removing a toxin;
• stopping smoking;
• blocking a molecular pathway.

100. Specificity is not universal

One cause can have many effects.

One effect can have many causes.

Therefore one-to-one specificity is not required in many complex biological systems.

101. Hill considerations are not a scoring algorithm

Checking five boxes does not prove causation.

Missing three boxes does not disprove it.

The considerations guide scientific judgment after alternative explanations have been examined.

---

102. Causation in school Science

Young students often use the word “because” before they have built a causal chain. School Science can develop causal reasoning gradually without introducing advanced causal-inference terminology too early.

Primary 3: observe before explaining

P3 pupils should first separate:

• what changed;
• what was observed;
• what they think caused the change.

Example:

“The paper became wet” is an observation.

“Water caused the paper to become wet” is a causal explanation.

The distinction matters because observations are evidence; causes are interpretations of why the evidence occurred.

Primary 4: one changed condition at a time

P4 pupils can learn the fair-test idea:

change one relevant condition → observe the outcome → keep other relevant conditions comparable

This is the classroom foundation of causal inference.

Primary 5: compare explanations

P5 pupils should begin asking:

• what else changed?
• could another factor explain the result?
• was the comparison fair?
• what evidence supports this causal claim?

This introduces confounding in age-appropriate form.

Primary 6: mechanism plus evidence

P6 pupils can build complete chains:

changed condition → affected process → evidence → observed consequence

They should also recognise when a question supports only association rather than cause.

Secondary Science: variables and causal design

Secondary students can work explicitly with:

• independent variables;
• dependent variables;
• controlled variables;
• confounders;
• repeated measurements;
• fair comparison.

The deeper lesson is not the vocabulary.

It is how design creates or weakens causal interpretation.

JC and advanced study: counterfactual causation

Older students can learn:

• randomisation;
• potential outcomes;
• causal diagrams;
• natural experiments;
• mediation;
• effect modification;
• identification assumptions.

The scientific question becomes:

What comparison approximates the world in which the exposure changed but other relevant causal conditions did not?

---

103. Causation Casebook 1 — Fertiliser and plant growth

Observation:

Plants given fertiliser are taller.

Possible explanations:

• fertiliser caused increased growth;
• fertilised plants also got more water;
• fertilised plants began taller;
• light conditions differed;
• measurement differed.

A stronger causal design randomises similar plants to fertiliser or control and keeps relevant conditions comparable.

Mechanistic evidence can then examine how nutrients alter plant processes.

104. Casebook 2 — Music and concentration

Students listening to music complete tasks faster.

Does music cause better concentration?

Possibilities:

• music helps;
• students who like music choose to use it;
• experienced students multitask more;
• task difficulty differs.

Random assignment to music versus silence can strengthen the causal contrast.

105. Casebook 3 — Tutoring and achievement

Students receiving tuition have higher scores.

Possible causal structures include:

• tuition improves learning;
• higher-income families purchase more tuition and other resources;
• struggling students seek more tuition, creating a negative selection effect;
• motivated families seek tuition and support study in other ways.

The raw association can therefore overestimate, underestimate or reverse the causal effect.

106. Casebook 4 — Exercise and health

People who exercise more often have better health.

Potential confounders:

• age;
• income;
• diet;
• baseline health;
• smoking;
• sleep.

Randomized exercise interventions can estimate some causal effects, but long-term adherence and generalisation remain challenges.

107. Casebook 5 — Coffee and heart disease

Suppose coffee drinkers show higher disease risk.

Before claiming causation, ask whether smoking, occupation, sleep or another lifestyle factor explains part of the association.

This is the confounder audit.

108. Casebook 6 — Shoe size and reading ability

In children, shoe size can correlate with reading ability.

Age causes both.

This is a simple demonstration that correlation can be real while causal direction between the measured variables is absent.

109. Casebook 7 — Ice cream and drowning

Ice-cream sales and drowning incidents can both rise in hot weather.

Temperature or season is a common cause.

Buying ice cream does not cause drowning.

110. Casebook 8 — Firefighters and fire damage

More firefighters are present at larger fires.

Larger fires also cause more damage.

Fire severity causes both firefighter count and damage.

If we ignored severity, firefighters could look harmful.

111. Casebook 9 — Rain and traffic accidents

Rain may causally affect accident risk through:

• road friction;
• visibility;
• driver behaviour.

But traffic volume also changes in rain.

A causal analysis should account for exposure to driving itself.

112. Casebook 10 — Air pollution and hospital admissions

Causal inference can use:

• time-series designs;
• natural experiments;
• geographic comparisons;
• mechanistic evidence;
• dose-response relationships.

Weather, seasonality and socioeconomic conditions can confound or modify effects.

113. Casebook 11 — Smoking and lung cancer

The causal conclusion rests on far more than one association.

Evidence includes:

• large epidemiologic associations;
• dose-response;
• temporality;
• replication;
• cessation effects;
• mechanistic evidence;
• consistency across populations.

This is a classic example of causal inference without randomising exposure.

114. Casebook 12 — Vaccination and disease

Randomized trials can establish efficacy before deployment.

Later observational studies can evaluate effectiveness in real populations.

Causal complications include:

• who chooses vaccination;
• prior immunity;
• age;
• exposure patterns;
• time since vaccination.

Design and adjustment must follow the causal question.

115. Casebook 13 — Medicine and adverse effects

Rare harms may not appear in pre-approval trials because sample sizes are limited.

Post-market observational data can identify signals.

Causal interpretation then asks:

• does risk rise after exposure?
• is there dose-response?
• does risk fall after stopping?
• could disease severity explain treatment choice?
• is there a mechanism?

116. Casebook 14 — School policy and test scores

A policy begins in some schools but not others.

Simple before-after comparison is weak because:

• scores may already be trending;
• schools may differ;
• other reforms may occur simultaneously.

Difference-in-differences or threshold-based designs may strengthen the causal contrast if assumptions are plausible.

117. Casebook 15 — Class size and learning

Smaller classes may improve outcomes through:

• more feedback;
• more student participation;
• less disruption.

But schools with smaller classes may also have different resources.

Random assignment or policy thresholds can help isolate class-size effects.

118. Casebook 16 — Retrieval practice and delayed learning

Suppose retrieval practice produces higher delayed test scores than rereading.

For causal interpretation, hold comparable:

• study time;
• content exposure;
• feedback;
• prior knowledge.

Then test whether the effect transfers to novel questions.

119. Casebook 17 — Screen time and mental health

An association can reflect:

• screen time affecting mental health;
• mental-health difficulty increasing screen use;
• social isolation affecting both;
• sleep mediating effects;
• different types of screen use having different effects.

The causal question must be narrower than “Are screens bad?”

120. Casebook 18 — Social media and political attitudes

Causal pathways can include:

• content exposure;
• social reinforcement;
• self-selection;
• algorithmic ranking;
• network structure.

Observational correlations between usage and attitudes are not enough to identify one pathway.

121. Casebook 19 — Job training and employment

People who enrol in training may differ in motivation or labour-market prospects.

Randomized programme access, lotteries or eligibility cut-offs can provide stronger causal leverage.

122. Casebook 20 — Minimum wage and employment

Policy changes create observational causal questions because jurisdictions cannot usually be randomised easily.

Researchers may use:

• difference-in-differences;
• border comparisons;
• synthetic controls;
• event studies.

Results depend on assumptions about comparable counterfactual trends.

123. Casebook 21 — Tax and consumption

If a tax is introduced, consumption may change.

But other conditions also move over time.

A causal design needs a credible comparison or model of the untreated trend.

124. Casebook 22 — Advertising and sales

Companies often advertise more when they expect high demand.

Simple correlation between ad spend and sales can therefore be confounded by expected demand.

Randomized geographic advertising experiments can estimate incremental causal lift.

125. Casebook 23 — Recommendation systems

A recommender shows products to users likely to click.

High click-through does not prove the recommendation caused the click.

Randomized holdouts can estimate incremental effect.

126. Casebook 24 — Price and demand

Observed price and demand are jointly determined.

When demand is high, sellers may raise prices.

This can create a positive price-demand association even if higher prices causally reduce demand.

Econometric causal inference uses instruments or experiments to separate the directions.

127. Casebook 25 — Traffic apps and congestion

A routing app may cause drivers to shift routes.

But users who choose the app may differ from non-users.

At scale, widespread use changes the traffic system itself.

The causal effect is therefore both individual and systemic.

128. Casebook 26 — Machine maintenance and failure

Machines receiving more maintenance may appear to fail more often because high-risk machines are maintained more.

This is confounding by indication again.

A preventive-maintenance causal effect needs comparable machines or randomized maintenance schedules where feasible.

129. Casebook 27 — Temperature and equipment failure

Higher temperature can causally accelerate degradation.

But load may raise both temperature and failure risk.

Load is a potential confounder.

Controlled experiments can isolate temperature more cleanly.

130. Casebook 28 — Software update and error rate

An error spike begins after a software update.

Temporal order supports suspicion.

But other causes include:

• traffic increase;
• hardware change;
• logging change;
• external service failure.

Rollback or staged deployment can provide a causal test.

131. Casebook 29 — Climate forcing and temperature

Causal attribution uses:

• physical mechanism;
• energy balance;
• observed fingerprints;
• model experiments;
• counterfactual simulations.

The causal question is not answered by correlation alone.

132. Casebook 30 — Deforestation and rainfall

Deforestation can affect rainfall through:

• evapotranspiration;
• surface energy;
• moisture recycling.

But deforestation also correlates with land use, fires and regional climate variability.

Mechanism, satellite observations and natural experiments can work together.

133. Casebook 31 — Species decline and climate

Species decline may reflect:

• temperature;
• habitat loss;
• disease;
• invasive species;
• harvesting.

Causal attribution requires separating concurrent pressures.

134. Casebook 32 — Predator removal and ecosystem change

Removing a predator can alter prey and vegetation.

This is a trophic cascade.

Strong causal evidence can come from:

• before-after comparisons;
• control sites;
• reintroduction;
• mechanistic population data.

135. Casebook 33 — Genetic variant and disease

A genetic variant associated with disease may:

• causally affect risk;
• be linked to the true causal variant;
• reflect population structure.

Functional experiments and fine mapping strengthen causal interpretation.

136. Casebook 34 — Gene knockout

Removing a gene and observing a changed phenotype can support causation.

But compensatory pathways, developmental effects and off-target changes must be considered.

137. Casebook 35 — Molecular pathway inhibition

If inhibiting Mediator M blocks the effect of X on Y, that supports a mediating role.

But inhibitors can have off-target effects.

Multiple independent interventions strengthen the mechanism claim.

138. Casebook 36 — Asteroid impact and extinction

Causal inference about historical events uses:

• timing;
• geological markers;
• impact evidence;
• global environmental consequences;
• fossil turnover.

Historical causation can be scientifically tested through traces.

139. Casebook 37 — Volcanoes and climate anomalies

Large eruptions inject aerosols that can alter radiative balance.

Temporal alignment, physical mechanism and spatial fingerprints strengthen causal attribution.

140. Casebook 38 — Solar activity and climate

Solar variation is a real climate forcing.

The causal question is how much of an observed pattern it explains relative to other forcings.

Competing causal contributions can coexist.

141. Casebook 39 — Intervention causes improvement but mechanism is wrong

An experiment can establish that Treatment X works while the proposed mechanism remains incorrect.

Causal effect and causal mechanism are related but distinct questions.

142. Casebook 40 — Mechanism is correct but effect is tiny

A real mechanism can produce a small population effect because:

• pathway contribution is minor;
• countervailing mechanisms exist;
• exposure variation is limited.

Mechanistic truth does not determine practical magnitude automatically.

---

143. Causal reasoning in experiments

Every controlled experiment encodes a causal question.

Ask:

• what is manipulated?
• what outcome is measured?
• what comparison estimates the counterfactual?
• what pathways could still differ?

This is why “fair test” is deeper than keeping random variables constant.

144. A control group is a counterfactual approximation

The control group stands in for:

“What would have happened to the treated group without treatment?”

The closer that approximation, the stronger the causal interpretation.

145. Random assignment separates treatment from prognosis

Without randomisation, treatment may be given because of expected outcome.

Random assignment breaks that selection path.

146. Random sampling and random assignment are different

Random sampling helps represent a population.

Random assignment helps estimate causal effects within the study.

A study can have one without the other.

147. Internal validity versus external validity

Internal validity: is the causal effect credible in the study?

External validity: does the effect transport to the target population?

Strong causation work needs both eventually.

148. Transportability

To transport a causal effect, ask:

• which effect modifiers differ?
• does treatment operate through the same mechanism?
• are outcome definitions comparable?
• are baseline risks different?

149. Generalisation is itself a causal problem

A treatment effect can vary by context.

Moving from study sample to real population therefore requires causal assumptions about effect modification and selection.

150. Causal inference does not end with one estimate

A mature causal programme asks:

• average effect;
• who benefits;
• who is harmed;
• mechanism;
• duration;
• transport;
• interaction;
• policy consequence.

“Does X cause Y?” is often only the first question.

---

151. Interference: one person’s treatment can affect another person’s outcome

Many simple causal frameworks assume no interference.

That means:

one unit’s outcome depends only on its own treatment, not other units’ treatments.

This can fail in:

• vaccination;
• education;
• social networks;
• infectious disease;
• traffic;
• markets.

152. Vaccination and spillovers

If one person is vaccinated, transmission risk to others can change.

The causal effect includes:

• direct protection;
• indirect protection;
• population-level transmission change.

Individual treatment effects and population effects are therefore different estimands.

153. Education interference

A classroom intervention may change:

• peer behaviour;
• teacher attention;
• group norms;
• collaboration.

A treated student can affect untreated classmates.

Cluster randomisation may be more appropriate than individual randomisation.

154. Network interference

In social networks, treatment can spread through:

• information;
• behaviour;
• social influence;
• resource sharing.

Causal effects depend on network position and treatment coverage.

155. Stable unit treatment value assumption

Simple potential-outcome notation often assumes:

• no relevant interference;
• well-defined treatment versions.

This collection of ideas is associated with the stable unit treatment value assumption, often abbreviated SUTVA.

When SUTVA fails, the causal question must be expanded.

156. Multiple versions of treatment

“Tutoring” is not one intervention.

It can differ in:

• group size;
• duration;
• teacher quality;
• curriculum;
• feedback;
• frequency.

If different versions have different effects, a vague treatment definition weakens causal inference.

157. Well-defined interventions

A causal question should make clear:

• what is changed;
• how it is changed;
• compared with what;
• for how long.

Otherwise potential outcomes are ambiguous.

---

158. Time-varying treatment

Many interventions change over time.

Examples:

• medication dosage;
• study intensity;
• exercise;
• policy exposure;
• machine maintenance.

The causal effect depends on the treatment history, not one binary label.

159. Dynamic treatment regimes

A dynamic regime says:

“Give Treatment A if condition C is met; otherwise give B.”

This resembles real decision-making.

Causal inference can compare regimes rather than fixed treatments.

160. Treatment-confounder feedback

Suppose:

• earlier treatment changes health status;
• health status changes future treatment;
• health status also predicts outcome.

Health status is both mediator and time-varying confounder.

Standard regression can fail.

161. Marginal structural models

Marginal structural models use inverse probability weighting to estimate causal effects under time-varying confounding affected by prior treatment.

The method illustrates a broader lesson:

causal structure determines the appropriate statistical method.

162. The g-formula

The g-formula computes expected outcomes under hypothetical treatment strategies using models for the observed data process.

It can estimate effects of dynamic interventions under causal assumptions.

163. G-estimation

G-estimation is another family of methods for structural nested models.

The technical details vary, but the causal goal is similar:

estimate counterfactual treatment effects in longitudinal settings.

---

164. Mediation asks how effects travel

Total effect:

How much does X change Y overall?

Indirect effect:

How much of that change operates through mediator M?

Direct effect:

How much remains through pathways not involving M?

These quantities require additional causal assumptions beyond estimating the total effect.

165. Natural direct and indirect effects

Some mediation frameworks define counterfactual quantities that combine treatment and mediator values from different hypothetical worlds.

These effects can be difficult to identify and interpret when mediator–outcome confounding is affected by treatment.

Mediation is scientifically powerful but assumption-heavy.

166. Interventional mediation effects

Alternative mediation definitions use hypothetical interventions on mediator distributions rather than cross-world counterfactuals.

These can sometimes align better with policy questions.

167. Mechanism versus mediation

A measured mediator can lie on a statistical pathway without being the full biological or physical mechanism.

Mechanism can contain multiple mediators and hidden processes.

Do not equate one mediation coefficient with complete mechanistic explanation.

---

168. Selection after treatment

Conditioning on a post-treatment variable can create bias.

Example:

analysing only people who survived treatment.

If treatment affects survival and survival is related to outcome, the selected groups are no longer comparable.

169. Truncation by death

Some outcomes are undefined when a person dies.

Example:

quality-of-life score at one year.

Comparing only survivors can be misleading because treatment changes who survives.

Causal inference must define the question carefully.

170. Principal stratification

Principal stratification defines groups based on potential values of a post-treatment variable under different treatments.

It can formalise some problems of post-treatment selection.

Interpretation remains complex because principal strata are partly unobserved.

171. Missing outcomes

Missing data can threaten causal inference if missingness depends on:

• treatment;
• outcome;
• prognosis;
• post-treatment variables.

Complete-case analysis may create selection bias.

172. Loss to follow-up

If participants with poor outcomes leave one group more often, observed outcomes among those remaining can be biased.

Use:

• retention;
• missingness models;
• weighting;
• sensitivity analysis.

173. Competing risks

One event can prevent another event from occurring.

Example:

death can prevent observation of disease recurrence.

The causal estimand must specify how competing events are handled.

---

174. Feedback loops

Causal systems often contain feedback.

Example:

• income affects education;
• education affects future income;
• income later affects further education.

Simple acyclic diagrams cannot represent simultaneous cycles directly without indexing variables by time.

175. Time-indexed causal diagrams

Feedback can be represented by separating:

• X at Time 1;
• Y at Time 2;
• X at Time 3.

The arrows remain forward in time.

176. Reciprocal causation

Two processes can influence one another over time.

Examples:

• sleep and stress;
• health and employment;
• traffic and route choice;
• belief and media exposure.

A cross-sectional correlation cannot separate the directions.

177. Simultaneity

In economics and other systems, variables can be jointly determined.

Price and quantity are a classic example.

Observed correlation can therefore reflect simultaneous equations rather than one simple directional effect.

---

178. Prediction is not causation

A model can predict Y accurately from X even if X does not cause Y.

Example:

postal code can predict many outcomes because it proxies social and environmental factors.

Changing the postal code label itself does not create the causal change.

179. Causation is not required for every decision

If the goal is:

“Which machine is most likely to fail tomorrow?”

a predictive model may be enough.

If the goal is:

“Which maintenance action will reduce failure?”

we need causal knowledge.

180. Intervention changes the data-generating process

A purely predictive association can break when we intervene on the predictor.

This is why causal models are needed for policy and action.

181. Proxy variables

A proxy can predict because it tracks the true cause.

Intervening on the proxy may do nothing.

Prediction asks:

“What forecasts Y?”

Causation asks:

“What would change Y?”

182. Anti-causal prediction

Sometimes the outcome causes the predictor.

Example:

disease changes a biomarker.

The biomarker can predict disease while not causing it.

This can be useful for diagnosis but not necessarily treatment.

183. Causal variables can be weak predictors

A true causal factor may explain only a small fraction of individual variation.

Example:

a treatment can reduce risk modestly even though many other factors determine outcome.

Causal importance and predictive importance are different dimensions.

---

184. Causal discovery

Causal-discovery algorithms try to infer aspects of causal structure from observational data under assumptions.

They may use:

• conditional independence;
• temporal information;
• non-Gaussian structure;
• invariance;
• interventions.

These methods can generate valuable hypotheses.

185. Causal discovery is assumption-dependent

Different causal structures can produce the same observed distribution.

Algorithms therefore require assumptions such as:

• causal sufficiency;
• faithfulness;
• acyclicity;
• functional-form restrictions.

Output should be interpreted as model-based inference, not automatic truth.

186. Conditional independence

If two variables become independent after conditioning on a third, that pattern can reveal causal structure under assumptions.

But measurement error or hidden variables can complicate interpretation.

187. Invariance across environments

A causal relationship may remain stable when non-causal correlations change across environments.

This idea motivates methods that search for invariant predictive relationships under interventions or domain shifts.

188. Interventional data are especially informative

Observational data tell us what happened naturally.

Interventions deliberately change one part of the system.

This can break equivalences among causal graphs.

189. Causal discovery and scientific mechanism should interact

An algorithm can suggest:

X → Y.

Domain science asks:

• is temporal order possible?
• is the mechanism plausible?
• can intervention test it?
• does independent evidence support it?

Computational causal discovery should generate tests, not close the case.

---

190. Heterogeneous causal effects

Average effects can hide important variation.

Examples:

• drug helps one genotype more;
• teaching method helps novices more than experts;
• policy affects regions differently.

Effect heterogeneity can be scientifically and ethically important.

191. Subgroup analysis

Subgroup effects should ideally be:

• pre-specified;
• mechanistically motivated;
• supported by interaction tests;
• replicated.

Searching many subgroups after results appear can produce false discoveries.

192. Individual treatment effects

People often ask:

“Will this treatment help me?”

Population causal effects do not directly reveal every individual’s counterfactual response.

Estimating individual effects is extremely difficult because the individual counterfactual remains unobserved.

193. Personalised causal inference

Personalised treatment aims to estimate effects conditional on patient characteristics.

Methods include:

• causal forests;
• meta-learners;
• Bayesian hierarchical models.

These need large data, careful validation and strong assumptions.

194. Average treatment effect can still guide policy

Even when individual effects vary, an average effect can be useful for population decisions.

The relevant estimand depends on the decision-maker.

195. Distributional causal effects

Treatment can change:

• mean;
• variance;
• tail risk;
• inequality;
• probability of threshold crossing.

Average outcome alone may hide important causal consequences.

196. Quantile treatment effects

Quantile effects ask how treatment changes different parts of the outcome distribution.

They can reveal heterogeneity that mean differences miss.

---

197. Transporting causal effects

A causal effect estimated in Study Population S may not equal the effect in Target Population T.

Reasons include:

• different baseline risk;
• different effect modifiers;
• different treatment implementation;
• different environment;
• different adherence.

198. Transport diagrams

Causal diagrams can include selection indicators showing where populations differ.

This helps reason about which variables must be measured to transport an effect.

199. Standardisation to target population

If effect modifiers are measured, study estimates can sometimes be reweighted or standardised to the target population distribution.

200. External validity is not an afterthought

A causal estimate can be internally excellent and practically irrelevant if it does not transport to the people or systems where the decision will be made.

Define the target population at the beginning.

---

201. Causal effect versus causal attribution

Effect question:

“What is the average effect of smoking on lung cancer risk?”

Attribution question:

“Did smoking cause this individual’s cancer?”

These are different problems.

Population effects do not identify individual causation with certainty.

202. Probability of causation

In some legal or epidemiologic settings, researchers discuss the probability that an exposure caused an individual’s outcome.

This requires stronger assumptions than estimating a population relative risk.

Individual attribution is difficult because multiple sufficient causal pathways can produce the same disease.

203. Causes of effects versus effects of causes

Effects of causes: What would changing X do to Y?

Causes of effects: Why did this particular Y occur?

Modern causal inference is generally stronger at effects-of-causes questions because interventions can be defined more clearly.

204. Root-cause analysis

Engineering asks:

“Why did this failure happen?”

Possible causes form a tree:

• design;
• material;
• maintenance;
• operator action;
• environment;
• software.

Root-cause analysis is a causal attribution problem.

205. Necessary condition versus contributory factor

One failure may require several conditions together.

Removing any one could prevent the outcome.

Causal responsibility can therefore be distributed across a sufficient-cause set.

206. Prevention paradox

A factor can cause many cases at population level while being a modest risk factor for each person.

Conversely, a strong individual risk factor can cause few population cases if exposure is rare.

Causal importance depends on both effect size and prevalence.

207. Population attributable fraction

Population attributable fraction estimates the proportion of outcomes that might be prevented if an exposure were removed under strong assumptions.

It combines causal effect with exposure prevalence.

Interpretation depends on whether elimination is well-defined and whether other factors remain unchanged.

208. Causal risk difference, ratio and odds

Causal effects can be expressed on different scales:

• risk difference;
• risk ratio;
• odds ratio;
• hazard ratio;
• mean difference.

The scientific interpretation changes with scale.

209. Absolute and relative effects

A 50% relative risk reduction can mean:

• risk falls from 20% to 10%;
• or from 2% to 1%.

The relative effect is the same.

The absolute effect is very different.

Decision-making needs both.

---

210. Causal Failure Atlas

Causal claims fail in recurring ways. A useful causal reasoner should recognise the pattern before reaching for a statistical fix.

Failure 1 — Correlation is called causation immediately

Pattern: X and Y move together, therefore X causes Y.

Repair: list reverse causation, confounders, selection and chance before choosing the causal interpretation.

Failure 2 — Reverse causation

Pattern: outcome influences the supposed cause.

Repair: establish temporal order and use longitudinal or interventional evidence.

Failure 3 — Common-cause confounding

Pattern: Z causes both X and Y.

Repair: randomise or control Z appropriately.

Failure 4 — Confounding by indication

Pattern: high-risk units are more likely to receive treatment.

Repair: emulate a target trial, adjust pre-treatment prognosis, or use randomisation/instruments.

Failure 5 — Adjusting for a mediator

Pattern: total-effect analysis controls away part of the causal pathway.

Repair: define whether the target is total, direct or indirect effect.

Failure 6 — Adjusting for a collider

Pattern: conditioning creates a spurious association.

Repair: draw the causal graph before adjustment.

Failure 7 — Selecting on survival

Pattern: only survivors are compared after treatment affects survival.

Repair: define the estimand carefully and use methods for truncation by death where appropriate.

Failure 8 — Immortal-time bias

Pattern: treated participants must survive long enough to be classified as treated.

Repair: align eligibility, treatment assignment and time zero.

Failure 9 — Time-varying confounding mishandled

Pattern: prior treatment changes a confounder that affects later treatment and outcome.

Repair: use longitudinal causal methods such as weighting or g-methods.

Failure 10 — Overcontrol

Pattern: every measured variable is included automatically.

Repair: adjust only according to causal structure and estimand.

Failure 11 — Residual confounding

Pattern: confounder is measured poorly or coarsely.

Repair: improve measurement and perform sensitivity analyses.

Failure 12 — Unmeasured confounding

Pattern: an important common cause is absent from the data.

Repair: seek natural experiments, instruments, negative controls or robustness analysis.

Failure 13 — Selection bias

Pattern: inclusion depends on exposure and outcome-related factors.

Repair: model the selection process or redesign sampling.

Failure 14 — Informative loss to follow-up

Pattern: dropout relates to treatment and outcome.

Repair: improve retention and use appropriate weighting/imputation with sensitivity analysis.

Failure 15 — Misclassification

Pattern: exposure or outcome is measured incorrectly.

Repair: validate measurement and model error where possible.

Failure 16 — Differential measurement error

Pattern: measurement quality differs by treatment or outcome status.

Repair: blind assessment and standardise instruments.

Failure 17 — Regression to the mean

Pattern: extreme cases improve naturally after being selected for extreme values.

Repair: use controls and repeated baseline measurements.

Failure 18 — Secular trend mistaken for intervention effect

Pattern: outcome was already changing before policy.

Repair: model pre-trends and use comparison groups.

Failure 19 — Simultaneous policy changes

Pattern: several interventions begin together.

Repair: find variation separating them or acknowledge underidentification.

Failure 20 — Wrong counterfactual

Pattern: comparison group does not represent what would have happened without treatment.

Repair: redefine design around the causal estimand.

Failure 21 — Inadequate positivity

Pattern: some covariate groups never receive one treatment.

Repair: restrict target population or use a better-supported causal contrast.

Failure 22 — Vague treatment versions

Pattern: “treatment” means many different interventions.

Repair: define treatment versions explicitly.

Failure 23 — Interference ignored

Pattern: one unit’s treatment affects others.

Repair: use cluster/network estimands or model spillovers.

Failure 24 — Wrong effect scale

Pattern: relative effect presented without absolute risk.

Repair: report causal effects on scales relevant to decisions.

Failure 25 — Average effect hides heterogeneity

Pattern: one subgroup benefits, another is harmed.

Repair: pre-specify plausible effect modifiers and validate subgroup effects.

Failure 26 — Post-hoc subgroup fishing

Pattern: many subgroups searched until one shows a large effect.

Repair: label exploratory and replicate prospectively.

Failure 27 — Mechanism asserted from treatment effect

Pattern: X changes Y, therefore pathway M must explain it.

Repair: measure or intervene on M separately.

Failure 28 — Treatment effect denied because mechanism unknown

Pattern: causal experiment succeeds but mechanism is incomplete.

Repair: separate effect estimation from mechanistic explanation.

Failure 29 — Statistical adjustment treated as magic

Pattern: a large regression is called causal because many covariates were included.

Repair: justify adjustment using causal assumptions.

Failure 30 — Predictive importance mistaken for causation

Pattern: feature strongly predicts outcome, therefore changing feature changes outcome.

Repair: distinguish proxy from intervention variable.

Failure 31 — Causal effect mistaken for prediction

Pattern: a true cause is expected to explain most individual variance.

Repair: separate average causal effect from predictive R².

Failure 32 — Individual attribution from population association

Pattern: group risk ratio used to declare exact cause for one individual.

Repair: acknowledge multiple causal pathways and attribution uncertainty.

Failure 33 — Natural experiment treated as random automatically

Pattern: policy variation is called quasi-random without checking assignment.

Repair: defend the exogeneity mechanism explicitly.

Failure 34 — Parallel-trends assumption ignored

Pattern: difference-in-differences used with divergent pre-trends.

Repair: examine pre-treatment dynamics and alternative controls.

Failure 35 — Weak instrument

Pattern: instrument barely changes exposure.

Repair: assess strength and robustness; weak instruments can amplify bias.

Failure 36 — Invalid instrument

Pattern: instrument affects outcome through another path.

Repair: justify exclusion restriction using science and sensitivity analysis.

Failure 37 — Causal graph treated as proof

Pattern: arrows are assumed true because they are drawn.

Repair: treat the DAG as explicit assumptions requiring evidence.

Failure 38 — Causal discovery treated as oracle

Pattern: algorithm output is presented as definitive causality.

Repair: inspect assumptions and seek interventions or external evidence.

Failure 39 — One Hill factor treated as decisive

Pattern: dose-response alone is called proof.

Repair: evaluate the complete evidence architecture and alternative explanations.

Failure 40 — Hill factors counted like points

Pattern: five of nine equals causation.

Repair: use them as considerations, not an algorithm.

---

211. Causal Practice Bank: 30 reusable reasoning prompts

Practice 1 — Name the association

Describe only what varies together before using causal language.

Practice 2 — Reverse the arrow

Ask whether Y could cause X.

Practice 3 — Find a common cause

Generate one Z that could affect both X and Y.

Practice 4 — Define the intervention

Turn “X causes Y” into “What would happen to Y if we changed X from A to B?”

Practice 5 — Define the counterfactual

What would have happened to the treated group without treatment?

Practice 6 — Build a control group

Choose the comparison that best approximates the missing counterfactual.

Practice 7 — Randomise

Explain why random assignment changes the causal interpretation.

Practice 8 — Confounder or mediator?

Place the variable in time relative to treatment and outcome.

Practice 9 — Collider check

Ask whether the variable is caused by both exposure and outcome-related factors.

Practice 10 — Selection check

Who entered the dataset, and why?

Practice 11 — Timeline check

Draw exposure, confounders, mediators and outcome along time.

Practice 12 — Target estimand

Specify population, intervention, comparator, outcome and horizon.

Practice 13 — Total versus direct effect

Decide whether mediation pathways belong inside or outside the target effect.

Practice 14 — Spillover check

Can one person’s treatment affect another person’s outcome?

Practice 15 — Treatment-version check

Does “treatment” mean the same intervention for everyone?

Practice 16 — Positivity check

Do all relevant groups have some chance of both treatments?

Practice 17 — Natural experiment audit

What assignment mechanism creates the quasi-random variation?

Practice 18 — Difference-in-differences audit

Were trends similar before intervention?

Practice 19 — Instrument audit

Does the instrument affect outcome only through treatment?

Practice 20 — Sensitivity analysis

How strong must hidden confounding be to reverse the conclusion?

Practice 21 — Negative control

Find an exposure or outcome that should not be causally affected.

Practice 22 — Mechanism test

What intermediate variable should change if the causal pathway is correct?

Practice 23 — Block the mechanism

If M mediates X→Y, what happens if M is prevented from changing?

Practice 24 — Heterogeneity

Which subgroup might experience a different causal effect, and why?

Practice 25 — Transport

Which effect modifiers differ between study and target population?

Practice 26 — Prediction versus intervention

Ask whether changing the predictor itself would change the outcome.

Practice 27 — Population versus individual

Separate average effect from attribution of one case.

Practice 28 — Absolute versus relative effect

Translate a risk ratio into absolute risk change.

Practice 29 — Failure diagnosis

If a randomized trial and observational study disagree, list possible reasons before choosing sides.

Practice 30 — Full causal cycle

Take one real question through causal graph, estimand, design, identification assumptions, estimate, sensitivity, mechanism and transport.

---

212. Twenty myths about causation

Myth 1 — Correlation never indicates causation

False. Causes often create correlations. The point is that correlation alone is insufficient.

Myth 2 — Correlation proves causation if it is large

False. Strong confounding can create large associations.

Myth 3 — Randomized trials are always perfect

False. They can suffer from non-adherence, missing data, measurement problems and limited generalisability.

Myth 4 — Observational studies cannot establish causation

False. Strong causal evidence can emerge from converging observational designs and mechanisms.

Myth 5 — Adjusting for more variables is always better

False. Adjusting for mediators or colliders can introduce bias.

Myth 6 — A confounder is any variable correlated with treatment

False. Confounding depends on causal structure and outcome relevance.

Myth 7 — A mediator should always be controlled

False. It depends on whether the target is total or direct effect.

Myth 8 — A causal diagram proves the arrows

False. DAGs make assumptions explicit.

Myth 9 — A p-value shows causality

False. It addresses chance under a statistical model, not causal direction or bias.

Myth 10 — Mechanistic plausibility proves causality

False. Plausible mechanisms need empirical support.

Myth 11 — No known mechanism means no cause

False. Causal relationships can be detected before mechanisms are understood.

Myth 12 — Dose-response is required

False. Real causal relationships can be thresholded or non-monotonic.

Myth 13 — One cause should have one effect

False. Causes can have many effects and effects can have many causes.

Myth 14 — The average treatment effect applies to everyone

False. Individual effects can differ.

Myth 15 — Predictive features are causal features

False. Proxies can predict without causing.

Myth 16 — If intervention changes outcome, mechanism is known

False. Effect and mechanism are separate questions.

Myth 17 — Natural experiments are automatically random

False. Assignment assumptions must be defended.

Myth 18 — Hill criteria are a checklist proving cause

False. They are considerations supporting causal judgment.

Myth 19 — Causation is one universal philosophical definition

False. Different fields use interventionist, counterfactual, mechanistic and other causal frameworks.

Myth 20 — Causal inference ends after one effect estimate

False. Mechanism, heterogeneity, transport, duration and policy consequences remain.

---

213. Causal reasoning rubric

Dimension	Weak	Developing	Strong
Association	jumps to cause	lists alternatives	maps causal structures explicitly
Time	order unclear	notes chronology	aligns time zero, treatment and outcome
Counterfactual	no comparator	informal comparison	defines estimand and credible comparison
Confounding	ignored	measured confounders listed	graph-guided control + sensitivity
Mechanism	asserted	plausible pathway	pathway tested with interventions/mediators
Bias	statistics assumed sufficient	some bias considered	selection, measurement and missingness audited
Transport	universal claim	target population named	effect modifiers and generalisation assessed
Decision	relative effect only	absolute effect noted	effect scale matched to action

---

214. Causal Design Workbook: turning ordinary claims into causal questions

The difference between a vague causal claim and a scientifically useful causal question is usually the quality of the counterfactual. The following transformations show how to move from everyday language toward testable intervention questions.

Transformation 1 — “Fertiliser makes plants grow”

Causal question: What is the average effect of assigning Fertiliser F versus no fertiliser on dry mass after 21 days in Species S under comparable light, water and soil conditions?

Key confounders in observational data: starting plant size, watering, light.

Best design: randomized assignment of similar plants.

Mechanism: nutrient availability affects growth processes.

Transformation 2 — “Tutoring raises grades”

Causal question: What is the effect of offering Programme T versus usual study support on delayed subject performance over one term?

Potential confounders: family resources, motivation, prior attainment, school support.

Possible design: randomized programme access or credible eligibility threshold.

Transformation 3 — “Exercise improves mood”

Causal question: What is the effect of a defined exercise programme compared with an attention-matched control on mood score after eight weeks?

Potential confounders: baseline health, sleep, social activity.

Mechanism possibilities: physiological change, mastery, social interaction.

Transformation 4 — “Coffee causes insomnia”

Causal question: What is the effect of a specified caffeine dose at Time T versus placebo on sleep-onset latency and total sleep time?

Randomised crossover design: each participant can receive caffeine and placebo on different nights where appropriate.

Transformation 5 — “Phones cause poor sleep”

Causal question: What is the effect of restricting evening phone use during a defined pre-sleep window on objectively measured sleep?

Alternative pathways: content, light exposure, bedtime delay, stress.

Transformation 6 — “Class size affects learning”

Causal question: What is the effect of reducing class size from A to B on delayed transfer performance, under comparable curriculum and instruction?

Mechanism candidates: feedback frequency, participation, classroom management.

Transformation 7 — “Homework causes achievement”

Causal question: What is the effect of assigning Homework Strategy H versus H0 on learning outcomes?

Warning: students receiving more homework may attend different schools or have different teachers. Raw homework–achievement correlation is confounded.

Transformation 8 — “More revision improves marks”

Causal question: What is the effect of increasing retrieval-practice dosage from D1 to D2 on delayed transfer, holding total study time comparable?

Possible nonlinearity: fatigue or diminishing returns.

Transformation 9 — “Sleep causes better exam performance”

Causal question: What is the effect of extending sleep opportunity by one hour for one week on standardised performance?

Challenge: adherence and behavioural compensation.

Transformation 10 — “Stress causes lower grades”

Causal question: What would happen to performance if stress were reduced by a defined intervention?

Issue: “stress” itself is not an intervention. Define the causal manipulation.

Transformation 11 — “Air pollution causes asthma attacks”

Causal question: What is the effect of reducing pollutant P from level A to B on acute asthma events over a defined period?

Natural experiments: regulation, plant closure, traffic changes.

Transformation 12 — “Noise causes poorer learning”

Causal question: What is the effect of reducing classroom noise by Intervention N on comprehension or retention?

Mechanism: attention capture, cognitive load.

Transformation 13 — “Teacher experience causes better outcomes”

Causal question: What is the effect of being taught by a teacher with Experience Level E versus E0, holding other school factors as comparable as possible?

Challenge: teacher assignment is rarely random.

Transformation 14 — “School choice improves results”

Causal question: What is the effect of attending School A versus the next-best alternative on later outcomes?

Possible design: admission lotteries can provide causal leverage.

Transformation 15 — “Vaccination prevents disease”

Causal question: What is the effect of assignment to vaccination versus control on disease risk over Horizon H?

Population effect: also include indirect effects through reduced transmission where relevant.

Transformation 16 — “A drug reduces mortality”

Causal question: What is the effect of Drug D compared with Standard Care S on 30-day mortality among eligible patients?

Trial complications: adherence, crossover, loss to follow-up.

Transformation 17 — “Screening saves lives”

Causal question: What is the effect of invitation to Screening Programme S on disease-specific and all-cause outcomes?

Bias risk: lead-time and length biases can distort observational comparisons among screened cases.

Transformation 18 — “Early diagnosis improves survival”

Warning: longer measured survival from diagnosis can occur without extending life if diagnosis simply occurs earlier.

Causal question: Does earlier detection plus effective intervention change mortality, not merely time since diagnosis?

Transformation 19 — “Maintenance prevents breakdowns”

Causal question: What is the effect of Maintenance Schedule M versus M0 on failure probability under comparable load?

Confounding by indication: high-risk machines may receive more maintenance.

Transformation 20 — “Higher temperature causes battery degradation”

Causal question: What is the effect of operating at Temperature T2 versus T1 on capacity retention, holding charge protocol comparable?

Controlled experiment: strong causal route.

Transformation 21 — “Advertising increases sales”

Causal question: What is the incremental effect of showing Campaign A on purchase probability?

Design: randomized user or geographic holdout.

Transformation 22 — “Discounts increase purchases”

Potential confounding: discounts may be offered precisely when demand is weak.

Causal design: randomised price experiment where lawful and ethical.

Transformation 23 — “Recommendation systems increase engagement”

Causal question: What is the effect of exposure to recommendation R relative to an alternative recommendation or no recommendation?

Issue: ranking algorithms select what is shown.

Transformation 24 — “Policy P reduced emissions”

Causal question: How would emissions have evolved without P?

Possible designs: synthetic control, difference-in-differences, threshold assignment.

Transformation 25 — “Protected areas increase wildlife populations”

Confounding: protected areas may be placed where populations were already healthier or human pressure lower.

Causal design: matched comparisons, boundary discontinuities, longitudinal models.

Transformation 26 — “Deforestation reduces rainfall”

Causal challenge: deforestation locations differ systematically.

Evidence: physical mechanism, atmospheric models, natural experiments, satellite observations.

Transformation 27 — “A gene causes disease”

Causal nuance: gene variants can alter risk, penetrance can be incomplete, and environmental factors can interact.

Evidence: segregation, functional experiments, population genetics.

Transformation 28 — “A biomarker causes disease”

Warning: biomarker may be a consequence or proxy.

Test: intervention on pathway or valid instrument such as genetic variation where assumptions hold.

Transformation 29 — “A social network causes behaviour”

Challenges: homophily, shared environment, peer influence.

Causal design: randomised exposure or natural network shocks where feasible.

Transformation 30 — “Media exposure changes beliefs”

Confounding: people choose media aligned with prior beliefs.

Causal design: randomised content exposure, encouragement designs, natural channel variation.

---

215. Causal Decision Manual

Decision 1 — Do we need causation or only prediction?

If the decision is passive forecasting, prediction may suffice. If the decision changes X to alter Y, causal inference is needed.

Decision 2 — Can we randomise?

If ethical, feasible and relevant, randomisation is often the cleanest route to causal effect estimation.

Decision 3 — If we cannot randomise, what creates exposure variation?

Look for policy rules, timing, thresholds, instruments, longitudinal variation or natural shocks.

Decision 4 — What is the counterfactual?

State what would have happened to the treated units under the alternative strategy.

Decision 5 — Which variables are pre-treatment causes?

These may be confounders requiring control.

Decision 6 — Which variables are post-treatment?

Do not adjust automatically. They may be mediators or colliders.

Decision 7 — Is there interference?

If treatment spills over, redefine the causal effect at network or cluster level.

Decision 8 — Is treatment well-defined?

If not, refine the intervention before estimating its effect.

Decision 9 — Is positivity plausible?

If some strata never receive one treatment, narrow the target population or intervention.

Decision 10 — Is the comparison contemporaneous?

Historical controls may differ because of secular trends.

Decision 11 — Are trends comparable?

For difference-in-differences, inspect pre-treatment trajectories.

Decision 12 — Does the instrument have a plausible exclusion restriction?

If not, do not use instrumental-variable language casually.

Decision 13 — Is missingness treatment-related?

If yes, complete-case analysis may be biased.

Decision 14 — Is effect heterogeneity likely?

Pre-specify key modifiers based on mechanism or policy relevance.

Decision 15 — Does the causal effect transport?

Compare effect modifiers between study and target population.

Decision 16 — Which scale matters?

Absolute risk, relative risk, mean difference and odds ratio answer different decision questions.

Decision 17 — Do we need mechanism?

If the goal is intervention optimisation or transport, mechanism may be crucial even after the total effect is known.

Decision 18 — How strong is hidden-confounding vulnerability?

Run sensitivity analyses rather than assuming measured adjustment solved everything.

Decision 19 — Can negative controls expose bias?

Use them where a plausible unaffected exposure or outcome exists.

Decision 20 — What evidence would reverse the causal conclusion?

A causal claim should remain revisable.

---

216. Causal Audit Notebook

A causal notebook can contain:

1. causal question;
2. target population;
3. treatment strategies;
4. comparator;
5. time zero;
6. follow-up;
7. outcome;
8. causal estimand;
9. DAG;
10. confounder set;
11. mediators;
12. selection variables;
13. interference assumptions;
14. identification assumptions;
15. estimator;
16. sensitivity analyses;
17. transport assumptions;
18. mechanism evidence;
19. decision scale;
20. revision history.

217. Why the causal notebook matters

Modern causal inference can become statistically complex.

The notebook keeps the scientific question visible.

If an analyst cannot state:

• what intervention is compared;
• what counterfactual is intended;
• which assumptions identify it;

then the causal analysis is not ready, no matter how advanced the software.

---

218. Causal Graph Practice: ten structures

Structure 1 — Direct cause

X → Y

Simple direct effect.

Structure 2 — Confounding

X ← Z → Y

Z opens a backdoor association between X and Y.

Structure 3 — Mediation

X → M → Y

M lies on the causal pathway.

Structure 4 — Collider

X → C ← Y

Conditioning on C can create association.

Structure 5 — Direct and indirect effects

X → Y and X → M → Y

Total effect includes both paths.

Structure 6 — Confounded mediator

A variable causes both M and Y. Mediation analysis must address mediator–outcome confounding.

Structure 7 — Treatment-confounder feedback

Earlier X affects L; L affects later X and Y. Longitudinal g-methods may be needed.

Structure 8 — Selection collider

X and Y both influence study participation S. Analysing only S=1 can bias association.

Structure 9 — Instrument

Z → X → Y

Z must not affect Y through other paths under the instrumental-variable assumptions.

Structure 10 — Effect modifier

The X→Y effect differs by Z. This is not represented only by one simple arrow; the causal effect is conditional on Z.

---

219. Causal Integrity Rules

Rule 1 — Do not hide the causal question inside the statistical model

State the intervention and counterfactual in plain language first.

Rule 2 — Do not use causal verbs for descriptive results casually

“Associated with”, “predicts”, “linked to” and “causes” are not interchangeable.

Rule 3 — Do not adjust blindly

Every adjustment variable should have a causal reason.

Rule 4 — Do not hide design limitations

Unmeasured confounding, selection and measurement assumptions should be visible.

Rule 5 — Preserve null results

Failed causal hypotheses are scientific information.

Rule 6 — Preserve intervention definitions

Changing treatment versions changes the causal question.

Rule 7 — Preserve target populations

Transport is not automatic.

Rule 8 — Preserve effect scales

Report absolute as well as relative effects where decision-relevant.

Rule 9 — Preserve mechanism uncertainty

A causal effect can be real while the mechanism remains incomplete.

Rule 10 — Preserve reversibility

A causal claim should remain open to better designs and new evidence.

---

220. Causal Glossary

Association

A statistical relationship in which variables occur or change together more often than expected under a reference model. Association alone does not determine causal direction.

Backdoor path

A non-causal path between exposure and outcome that begins with an arrow into the exposure and can create confounding.

Bias

A systematic distortion between the estimated relationship and the causal quantity of interest.

Collider

A variable caused by two or more other variables. Conditioning on a collider can create spurious associations among its causes.

Confounder

A variable involved in a non-causal path that distorts the exposure–outcome association, typically because it influences both.

Confounding by indication

Bias created when treatment is preferentially given to people or units with different prognosis.

Consistency

An assumption linking observed outcomes to the potential outcome under the treatment actually received, requiring well-defined intervention versions.

Counterfactual

An outcome that would have occurred under an alternative intervention or condition not actually observed for the same unit at that time.

Causal diagram

A graphical representation of assumed causal relationships, often using directed arrows.

Causal discovery

Methods that infer aspects of causal structure from data under assumptions about independence, functional form, interventions or invariance.

Causal effect

The contrast between outcomes under different interventions or treatment strategies.

Causal estimand

The precise causal quantity being targeted, defined by population, intervention, comparator, outcome and time horizon.

Causal inference

The reasoning and methods used to estimate or judge cause–effect relationships from experiments, observations and assumptions.

Causal mechanism

The process through which a cause produces an effect.

Causal pathway

A sequence of causal links connecting exposure to outcome.

Causal attribution

The task of judging whether a cause contributed to a particular observed effect, distinct from estimating average effects of interventions.

Conditional exchangeability

An observational assumption that, after conditioning on sufficient measured covariates, treatment groups are comparable with respect to potential outcomes.

Difference-in-differences

A quasi-experimental design comparing changes over time in treated and comparison groups under a parallel-trends assumption.

Direct effect

A causal effect not transmitted through a specified mediator, under the chosen definition.

Effect modification

Variation in a causal effect across levels of another variable.

Exchangeability

The condition that compared groups are comparable with respect to the relevant potential outcomes.

External validity

The degree to which a causal effect can generalise or transport beyond the study sample.

G-computation

A method for estimating counterfactual outcomes under interventions using models for the observed data-generating process.

Identification

The step of showing that a causal estimand can be expressed in terms of the observed data distribution under stated assumptions.

Indirect effect

The portion of a causal effect transmitted through a specified mediator, under a defined mediation framework.

Instrumental variable

A variable that changes exposure and, under strong assumptions, affects outcome only through that exposure and is independent of relevant unmeasured causes.

Intention-to-treat effect

The causal effect of assignment to an intervention strategy, regardless of adherence.

Interference

A situation in which one unit’s treatment affects another unit’s outcome.

Internal validity

The credibility of the causal inference within the study population and design.

Mediator

A post-exposure variable on a causal pathway between exposure and outcome.

Natural experiment

A situation where external circumstances create treatment variation that can approximate random or otherwise exogenous assignment.

Negative control

An exposure or outcome expected not to be causally affected in the hypothesised way, used to detect bias.

Observational study

A study in which researchers observe exposure or treatment rather than assigning it randomly.

Per-protocol effect

The causal effect associated with following a defined treatment protocol as specified.

Positivity

The requirement that each relevant covariate pattern has a nonzero probability of receiving each treatment under comparison.

Potential outcome

The outcome a unit would have under a specific treatment or intervention.

Propensity score

The probability of treatment given measured covariates, often used for matching, weighting or adjustment.

Random assignment

Assignment of treatment by a random mechanism, strengthening causal inference by breaking systematic baseline links between treatment and prognosis.

Random sampling

Sampling units randomly from a population to support representativeness. It is distinct from random assignment.

Regression discontinuity

A quasi-experimental design exploiting a treatment rule that changes at a threshold.

Selection bias

Bias arising because inclusion in the analysed sample depends on variables related to exposure and outcome.

Sensitivity analysis

Analysis examining how causal conclusions change under alternative assumptions, including possible unmeasured confounding.

Spillover

An effect of one unit’s treatment on another unit’s outcome.

Target trial

The hypothetical ideal randomized trial that an observational analysis attempts to emulate.

Total effect

The overall causal effect through all pathways from exposure to outcome.

Transportability

The ability to extend a causal effect estimate from the study population to another target population under assumptions.

Type of treatment version

A particular way an intervention is implemented; multiple versions can matter when treatment is vaguely defined.

---

221. Advanced FAQ: Scientific Causation

1. Does correlation ever prove causation?

No single correlation, by itself, proves a causal relationship. But causal effects often create correlations, and correlations can contribute to a broader causal case when design, timing, mechanism and alternatives are considered.

2. Is randomisation the only way to establish causation?

No. Randomisation is especially powerful because it reduces confounding by design, but strong causal knowledge can also come from natural experiments, longitudinal evidence, mechanistic studies and converging observational designs.

3. Why is randomisation so powerful?

Because assignment is not determined by participants’ prognosis or pre-treatment characteristics. This makes treatment groups exchangeable in expectation before intervention.

4. Can a randomized trial still be biased?

Yes. Non-adherence, dropout, unblinded measurement, crossover and missing outcomes can introduce bias after randomisation.

5. Can an observational study ever be stronger than a small poor trial?

Yes. Study quality matters. A large carefully designed natural experiment can sometimes provide stronger evidence than an underpowered, poorly executed trial.

6. Does temporal order prove causation?

No. A cause must precede its effect, but many non-causal variables also precede outcomes.

7. Does dose-response prove causation?

No. Dose-response can strengthen a causal interpretation, but confounding can also produce gradients.

8. Does mechanism prove causation?

No. Mechanistic plausibility strengthens causal claims, but plausible mechanisms can be wrong.

9. Can causation exist without a known mechanism?

Yes. Causal effects can be established before the mechanism is fully understood.

10. What is the simplest definition of a causal effect?

The difference between outcomes under two different interventions or treatment strategies for the same target population.

11. Why can’t we observe an individual’s causal effect directly?

Because the same individual cannot simultaneously experience both treatment and no treatment at the same moment under identical conditions.

12. What is the counterfactual problem?

The missing outcome under the alternative treatment that was not received.

13. What is the purpose of a control group?

To approximate what would have happened to the treated group under the alternative condition.

14. Why not control for every variable?

Because adjusting for mediators or colliders can bias the causal estimate. Adjustment should follow the causal graph and estimand.

15. What is a confounder in plain language?

A factor that helps create a misleading treatment–outcome association because it influences both the treatment or exposure and the outcome.

16. What is a mediator in plain language?

A step in the pathway through which the cause produces the effect.

17. What is a collider in plain language?

A variable that is influenced by two other variables. Selecting or adjusting on it can make those causes appear associated.

18. Why do causal diagrams help?

They force researchers to state assumptions about which variables cause which others, making adjustment choices inspectable.

19. Can a DAG be tested from data?

Some implications can be tested, but many different graphs can fit the same data. DAGs need domain knowledge and experimental evidence.

20. Is an instrumental variable a magic substitute for randomisation?

No. Instruments rely on strong assumptions about how they affect treatment and outcome.

21. Is matching equivalent to randomisation?

No. Matching balances measured covariates; randomisation balances measured and unmeasured pre-treatment causes in expectation.

22. Do propensity scores solve unmeasured confounding?

No. They use measured covariates.

23. What is the role of sensitivity analysis?

To show how strong assumption violations, such as unmeasured confounding, would need to be to change the conclusion.

24. What is causal heterogeneity?

Variation in treatment effect across people, groups, contexts or times.

25. Can an intervention help some and harm others?

Yes. Average effects can conceal opposite subgroup effects.

26. How should subgroup effects be studied?

Use pre-specified, mechanism-informed modifiers and replicate the interaction rather than searching many subgroups after the fact.

27. What is transportability?

The question of whether a causal effect estimated in one study applies to another target population.

28. Why can a causal effect change by population?

Effect modifiers, baseline risks, treatment versions and environmental conditions can differ.

29. Is a cause always necessary?

No. Many causes increase risk without being required for the outcome.

30. Is a cause always sufficient?

No. Causes often operate as parts of larger sufficient sets.

31. Can several causes act together?

Yes. Interaction is common in biology, behaviour, engineering and social systems.

32. What is reverse causation?

The outcome influences the supposed cause rather than the proposed direction.

33. Can prediction be excellent without causal understanding?

Yes. A proxy can predict well even if changing the proxy does not change the outcome.

34. Can causal understanding be strong while prediction is weak?

Yes. A treatment can have a genuine average causal effect while individual outcomes remain highly variable.

35. Why is causal inference important for policy?

Policy changes conditions. To estimate the consequences of action, we need intervention effects, not only associations.

36. Why is causal inference important for medicine?

Treatment decisions require knowing what changes outcomes, not merely what predicts them.

37. Why is causal inference important for education?

Schools need to know whether interventions improve learning, not merely whether successful students tend to use them.

38. Why is causal inference important for engineering?

Root-cause analysis and intervention require distinguishing the factor that produces failure from conditions that merely accompany it.

39. Can causation be probabilistic?

Yes. A cause can increase the probability of an outcome without determining it in every case.

40. What is the final causal question?

Not simply “Are X and Y related?” but “What would happen to Y if we intervened on X in a defined way for this population under these conditions?”

---

222. Causation and ethics

Causal experiments can create ethical problems precisely because they intervene.

Researchers cannot ethically randomize people to:

• smoking;
• toxic pollution;
• serious deprivation;
• known harmful treatments.

Ethical constraints shape which causal designs are possible.

223. Equipoise

Clinical randomisation is ethically acceptable when genuine uncertainty exists about which treatment is better and other ethical requirements are met.

Once overwhelming evidence shows one treatment is superior, continued randomisation may become unethical.

224. Informed consent

Participants should understand:

• intervention;
• risks;
• alternatives;
• data use;
• right to withdraw.

Ethics and causal validity can interact because coercion or differential dropout can also bias results.

225. Causal questions without randomisation

When harmful exposure cannot be assigned, science relies on:

• natural experiments;
• observational studies;
• mechanistic evidence;
• historical events;
• policy changes.

Ethical science does not abandon causation; it changes design.

226. Fairness in causal interventions

An intervention may have different effects across groups.

Ethical evaluation should consider:

• who benefits;
• who is harmed;
• who bears costs;
• whether effect heterogeneity is measured.

Average benefit can hide unequal causal consequences.

227. Policy feedback

Policies change behaviour.

Behaviour changes later exposure and outcomes.

Causal evaluation must sometimes model:

• adaptation;
• spillovers;
• market responses;
• network effects.

A static treatment effect can miss system-level consequences.

228. General equilibrium effects

In economics and large social systems, changing one policy can alter prices, wages, migration and incentives.

The causal effect at small scale may differ from the effect when everyone receives the intervention.

Scale-up is a causal question, not a simple multiplication.

---

229. Causal Evidence Standards: no single universal hierarchy

People often ask for a simple hierarchy of causal evidence.

Randomized controlled trials are powerful for many intervention questions because random assignment reduces confounding by design.

But no single hierarchy works for every scientific domain.

Consider:

• harmful exposures cannot always be randomized;
• rare adverse effects need large observational data;
• long-latency outcomes may require decades of follow-up;
• geological and astronomical causes are studied through traces;
• system-wide policies may be impossible to randomize.

The stronger question is:

Which design creates the most credible counterfactual for this causal question, under ethical and practical constraints?

230. Evidence strength depends on design fit

A randomized trial is not automatically superior if:

• the intervention differs from the real policy;
• follow-up is too short;
• adherence is poor;
• sample is unrepresentative;
• outcome is measured badly.

A well-designed observational study can sometimes answer the practical causal question better.

231. Triangulation across designs

One of the strongest causal strategies is triangulation.

Use methods with different biases:

• randomized trial;
• cohort study;
• natural experiment;
• instrumental variable;
• mechanistic experiment;
• negative controls;
• time-series evidence.

If these independent approaches converge, shared bias becomes less likely.

232. Replication of causal effects

A causal result becomes stronger when it replicates across:

• populations;
• sites;
• research groups;
• treatment versions;
• measurement methods.

Exact numerical equality is not expected if effect modifiers differ.

233. Heterogeneous replication

Suppose Treatment A reduces risk by 20% in one population and 10% in another.

This is not necessarily replication failure.

The effect may vary because:

• baseline risk differs;
• adherence differs;
• genetics differ;
• co-interventions differ;
• treatment implementation differs.

Replication should evaluate compatibility with a causal mechanism, not demand one universal number.

234. Meta-analysis of causal studies

Meta-analysis combines estimates across studies.

It can improve precision and reveal heterogeneity.

But causal interpretation depends on the quality of the studies being combined.

Combining ten biased studies does not magically remove bias.

235. Study quality matters more than count

Ask:

• Were the studies targeting the same estimand?
• Were treatments comparable?
• Were populations similar enough?
• Were confounders handled credibly?
• Were outcomes measured consistently?

Meta-analysis should synthesize causal questions, not just effect-size columns.

236. Publication bias in causal research

Positive or dramatic causal findings may be more likely to appear publicly.

This can inflate apparent evidence.

Study registration, reporting standards and publication of null results help reduce this distortion.

237. Registered causal analyses

Preregistration can specify:

• causal question;
• target estimand;
• DAG;
• adjustment set;
• primary outcome;
• analysis method;
• sensitivity analyses.

This makes causal reasoning auditable before outcomes tempt researchers to change the plan.

238. Causal reproducibility

To reproduce a causal analysis, another analyst should know:

• how treatment was defined;
• how time zero was set;
• how confounders were measured;
• which observations were excluded;
• which estimator was used;
• which weights or matches were created;
• how missing data were handled.

A coefficient without the design history is incomplete.

---

239. Causation in historical sciences

Some causes happened once or cannot be manipulated now.

Examples include:

• asteroid impacts;
• mass extinctions;
• continental movement;
• ancient climate changes;
• stellar events.

Historical sciences infer causation from traces.

240. Traces support past causal claims

A historical causal hypothesis predicts that certain evidence should remain:

• geochemical markers;
• fossil patterns;
• crater geometry;
• isotope ratios;
• sediment layers;
• orbital signatures.

If those traces are found independently, the causal claim gains support.

241. Retrodiction

Retrodiction asks:

“If this past causal process occurred, what should we observe now?”

This is scientifically analogous to prediction, except the target event lies in the past and the evidence remains in present traces.

242. Historical causation can use multiple counterfactual models

Researchers can compare:

• world with Event E;
• world without E;

through simulations, comparative cases and geological records.

The counterfactual is model-based rather than experimentally observed.

243. Historical sciences can achieve strong causal confidence

Lack of laboratory randomisation does not imply weak science.

Strong causal confidence can arise from:

• temporal sequence;
• mechanism;
• multiple independent traces;
• rival-hypothesis failure;
• quantitative modelling.

---

244. Causation in complex systems

Complex systems have:

• many causes;
• feedback loops;
• thresholds;
• adaptation;
• emergent behaviour.

A single-arrow causal story can be misleading.

245. Nonlinear effects

Doubling a cause does not always double the effect.

Causal relationships can be:

• thresholded;
• saturating;
• U-shaped;
• hysteretic;
• discontinuous.

Estimate the response function, not only one average contrast.

246. Threshold causation

A factor can have little effect until a threshold is crossed.

Examples:

• material stress;
• toxin exposure;
• ecosystem collapse;
• epidemic reproduction.

Thresholds create causal regimes.

247. Hysteresis

In some systems, returning the cause to its earlier value does not immediately restore the old state.

The system’s history matters.

Examples can occur in:

• magnetic systems;
• ecological regime shifts;
• social systems;
• materials.

Causal models need state dependence.

248. Feedback amplification

A small causal change can trigger a loop:

X increases Y → Y increases X → larger total change

Feedback can make the total causal effect much larger than the direct effect.

249. Negative feedback

Negative feedback can damp effects.

Example:

temperature rises → cooling process strengthens → temperature increase slows.

Causal response can be self-limiting.

250. Emergent causation

At system scale, causal relationships can emerge that are not meaningful for one component.

Examples:

• traffic congestion;
• market prices;
• ecosystem resilience;
• collective neural activity.

The causal variable may exist only at the higher level.

251. Downward causation language should be used carefully

Scientists sometimes say system-level states constrain component behaviour.

The phrase “downward causation” is philosophically debated.

Operationally, ask which measurable system variables change component outcomes.

Keep the claim tied to interventions and mechanisms.

252. Causal coarse-graining

A complex micro-system can sometimes be represented by a smaller set of macro variables that preserve the causal relationships relevant to the question.

This is a form of causal simplification.

The right level depends on the intervention.

---

253. Causation and AI

AI systems can support causal science by helping with:

• literature synthesis;
• DAG generation;
• hypothesis generation;
• heterogeneous-effect modelling;
• causal discovery;
• simulation.

But causal reasoning cannot be delegated blindly.

254. Language models can generate plausible causal stories

A fluent causal explanation is not evidence.

AI can propose:

“X causes Y through M.”

Science must still ask:

• was X before Y?
• could Z explain both?
• can M be measured?
• does intervention on X change Y?
• does blocking M alter the effect?

255. AI can amplify confounding mistakes

If training data contain observational associations, a model may repeat causal language found in those sources.

Researchers should check whether causal claims trace back to valid study designs.

256. Machine learning for nuisance functions

Modern causal estimators can use machine learning to estimate:

• propensity scores;
• outcome models;
• conditional effects.

Flexible prediction can improve causal estimation when combined with valid identification assumptions.

257. Double machine learning

Double/debiased machine learning uses sample splitting and orthogonalisation to estimate causal parameters while flexibly modelling nuisance functions.

The technical approach helps reduce some estimation bias.

It still depends on causal assumptions about confounding and treatment assignment.

258. Causal forests

Causal forests estimate heterogeneous treatment effects using tree-based methods.

They can explore which subgroups respond differently.

Subgroup findings should be validated independently.

259. AI cannot create the missing counterfactual automatically

More computation does not let us observe both potential outcomes for the same unit.

Identification still comes from:

• design;
• assumptions;
• interventions;
• natural variation.

260. AI-generated causal graphs need audit

An AI can propose a DAG from text or data.

Researchers should verify:

• temporal order;
• domain plausibility;
• missing variables;
• feedback;
• selection.

The graph is a hypothesis about structure.

---

261. Causal Reporting Standard

A strong causal report should state:

1. the causal question;
2. target population;
3. treatment and comparator;
4. time zero;
5. follow-up;
6. outcome;
7. estimand;
8. study design;
9. causal assumptions;
10. adjustment variables and why;
11. estimation method;
12. effect estimate on meaningful scales;
13. uncertainty interval;
14. sensitivity analysis;
15. mechanistic support;
16. transport limitations.

262. Report association when causation is not identified

If the design does not support a causal interpretation, use honest language:

• associated with;
• correlated with;
• predicts;
• linked to.

Scientific precision includes linguistic precision.

263. Report the assumptions

Causal estimates are conditional on assumptions.

Do not bury them in supplementary material if they are central to validity.

264. Report absolute effects

Where relevant, translate relative effects into absolute numbers.

Example:

“Risk fell from 10% to 7%, an absolute reduction of 3 percentage points and relative reduction of 30%.”

Both views matter.

265. Report heterogeneity carefully

Subgroup effects should include:

• pre-specification status;
• interaction evidence;
• uncertainty;
• replication status.

266. Report causal diagrams when useful

A DAG can make adjustment assumptions transparent to readers.

It also makes disagreement productive: critics can challenge a specific arrow or missing variable.

---

267. Causal Replication Standard

Replicating a causal claim should preserve the causal question while allowing a new setting.

Record:

• same treatment contrast?
• same outcome?
• same horizon?
• same population?
• same mechanism?
• same confounder structure?

Differences can reveal effect modification rather than simple failure.

268. Direct causal replication

Repeat the same design and estimand in a new sample.

This tests repeatability under similar conditions.

269. Conceptual causal replication

Use another design to estimate the same causal relationship.

Example:

• randomized trial;
• natural experiment;
• mechanistic intervention.

Convergence across designs is especially persuasive.

270. Mechanism replication

Test the proposed pathway independently.

If the total effect replicates but mechanism does not, the causal claim may survive while the explanation changes.

271. Transport replication

Estimate the effect in a different population.

This tests generalisability and effect modification.

272. Causal failure should update, not vanish

If a replication fails:

• preserve the result;
• compare design;
• check measurement;
• check treatment versions;
• test effect modifiers;
• update causal confidence.

---

273. Causal scope and boundaries

This guide focuses on what changes what.

Related scientific-method topics are treated separately:

• Evidence focuses on how observations change confidence.
• Experiments focus on how fair tests are built.
• Explanations focus on mechanisms.
• Hypotheses focus on candidate testable claims.
• Prediction focuses on forecasting unseen outcomes.
• Theories focus on broad explanatory frameworks.
• Uncertainty focuses on ranges, unknowns and confidence.

Causation connects those topics through the counterfactual question:

What would happen to Y if we changed X?

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

• How Scientific Hypotheses Work — candidate causal and non-causal claims.
• How Scientific Theories Work — broad explanatory frameworks.
• How Scientific Models Work — representations of systems and pathways.
• How Scientific Measurement Works — reliable exposure and outcome measurement.
• How Scientific Evidence Works — observations and confidence.
• How Scientific Experiments Work — intervention and test design.
• How Scientific Explanations Work — causal mechanisms and explanatory chains.
• How Scientific Prediction Works — association and model used for forecasting.
• How Scientific Uncertainty Works — uncertainty in causal estimates and assumptions.
• When Science Results Disagree — conflicting causal results and the next test.

275. Authoritative further reading

• CDC Field Epidemiology Manual — Analysing and Interpreting Data
• National Academies — Reference Manual on Scientific Evidence: Epidemiology and Causation
• National Academies — Reference Manual on Scientific Evidence: Statistical Evidence and Causal Inference
• PMC — Causal Inference and Confounding: A Primer for Interpreting and Conducting Infectious Disease Research

276. Causal Integrity Close

The scientific use of causal language carries a responsibility.

“Causes” is stronger than “is associated with”.

It implies that intervention matters.

It implies a counterfactual.

It implies that alternative explanations have been considered seriously.

A causal claim should therefore leave an audit trail:

• what was changed;
• what was compared;
• which confounders were addressed;
• which assumptions were made;
• which mechanisms support the pathway;
• which populations the claim applies to;
• what evidence would make us revise it.

277. How Scientific Causation Works: the larger lesson

Science does not earn causal knowledge by noticing that two variables move together.

It earns causal knowledge by building a credible comparison between worlds:

the world where X is changed and the world where it is not.

Sometimes randomisation creates that comparison.

Sometimes natural experiments do.

Sometimes longitudinal observation, mechanism, triangulation and historical evidence work together.

The methods differ because the systems differ.

The causal discipline remains the same:

define the intervention → define the counterfactual → map alternative paths → build a credible comparison → estimate the effect → test the mechanism → examine heterogeneity → transport carefully → revise when evidence changes.

Prediction asks what will happen.

Explanation asks why it happened.

Causation asks the action question:

What would change if we changed this?

That question sits at the heart of experiments, medicine, engineering, policy and scientific intervention.

---

278. The Causal Research Programme

A mature causal question rarely ends with one study. Strong causal knowledge grows through a programme in which different designs answer different parts of the problem.

Stage 1 — Detect the association

Begin by establishing whether exposure and outcome are related at all.

At this stage, use careful descriptive language.

Ask:

• how strong is the association?
• does it vary over time?
• does it replicate?
• what obvious biases could explain it?

Stage 2 — Establish temporal order

Determine whether the proposed cause precedes the outcome.

Use:

• longitudinal follow-up;
• historical records;
• exposure timing;
• time-stamped interventions.

Stage 3 — Map plausible confounders

Draw the causal graph before modelling.

Ask which pre-treatment variables could influence both exposure and outcome.

Stage 4 — Seek exogenous variation

Where possible, use:

• randomisation;
• policy thresholds;
• lotteries;
• natural shocks;
• instruments.

Exogenous variation strengthens the counterfactual comparison.

Stage 5 — Estimate total causal effect

Define the estimand clearly.

Estimate the effect of intervention versus comparator in the target population.

Stage 6 — Test mechanism

Measure intermediates.

Intervene on the pathway where feasible.

Ask whether the mechanism predicts additional observations.

Stage 7 — Test effect heterogeneity

Ask whether the effect differs by:

• age;
• genotype;
• baseline risk;
• environment;
• dose;
• implementation.

Stage 8 — Transport

Move the causal claim to new populations and settings.

Measure relevant effect modifiers.

Stage 9 — Scale up

Ask whether individual-level effects persist when the intervention changes the whole system.

Spillovers and equilibrium responses may appear.

Stage 10 — Monitor unintended effects

A causal intervention can produce:

• benefits;
• harms;
• adaptation;
• behaviour change;
• substitution effects.

The causal programme should follow all important outcomes, not only the intended one.

Stage 11 — Replicate with different designs

A randomized result plus a natural experiment plus mechanism evidence is stronger than repeated reliance on one method.

Stage 12 — Update policy or theory

Causal knowledge should change action and scientific understanding proportionally to the evidence.

---

279. Causation by field: physics

Physical sciences often have strong intervention and mechanism control.

Causal reasoning may involve:

• controlled changes in force;
• energy input;
• temperature;
• field strength;
• material properties.

Precise mathematical models can connect intervention to outcome.

280. Causation by field: chemistry

Chemical causation can be tested through:

• concentration changes;
• temperature changes;
• catalyst manipulation;
• isotopic tracing;
• reaction-pathway inhibition.

Mechanistic causal chains can often be studied directly.

281. Causation by field: biology

Biology is often multicausal and heterogeneous.

Experiments can manipulate:

• genes;
• proteins;
• environment;
• organisms;
• populations.

Redundancy and feedback mean one cause may have context-dependent effects.

282. Causation by field: medicine

Medicine combines:

• randomized trials;
• observational effectiveness studies;
• pharmacology;
• mechanistic biomarkers;
• post-market surveillance.

Causal knowledge must address both benefit and harm.

283. Causation by field: epidemiology

Epidemiology often studies exposures that cannot be randomized.

Strong causal work uses:

• cohort studies;
• case-control studies;
• natural experiments;
• negative controls;
• triangulation;
• causal diagrams.

284. Causation by field: ecology

Ecological systems contain:

• multiple species;
• feedback loops;
• spatial variation;
• rare disturbances.

Field experiments, manipulations and natural experiments complement long-term observation.

285. Causation by field: climate science

Climate attribution uses:

• physical mechanism;
• observational fingerprints;
• model experiments with and without forcings;
• historical records;
• energy-balance constraints.

The counterfactual can be simulated rather than literally observed.

286. Causation by field: geology

Geological causes are inferred through:

• stratigraphy;
• dating;
• geochemistry;
• tectonic models;
• natural comparisons.

Historical traces act as causal evidence.

287. Causation by field: astronomy

Astronomers cannot manipulate stars and galaxies directly.

Causal inference uses:

• natural variation;
• temporal sequence;
• physical theory;
• spectroscopy;
• comparative populations;
• simulations.

288. Causation by field: education

Education causal questions include:

• instructional methods;
• class size;
• feedback;
• curriculum;
• technology.

Randomized trials can be valuable, but clustering, implementation fidelity and spillovers matter.

289. Causation by field: economics

Economics frequently uses:

• randomized field experiments;
• instrumental variables;
• regression discontinuity;
• difference-in-differences;
• synthetic controls.

General-equilibrium responses can make small-scale effects differ from large-scale policy effects.

290. Causation by field: social science

Human behaviour creates:

• selection;
• feedback;
• adaptation;
• network effects;
• measurement challenges.

Causal claims therefore require careful design and transparent assumptions.

291. Causation by field: engineering

Engineering causation often focuses on:

• failure;
• load;
• material degradation;
• control systems;
• maintenance.

Experiments, simulation and root-cause analysis work together.

292. Causation by field: computer systems

Software and distributed systems can be manipulated precisely.

Causal diagnosis may use:

• A/B tests;
• feature flags;
• rollback;
• fault injection;
• counterfactual logging.

But user selection and network effects still matter.

---

293. Thirty final causal audit questions

1. What exactly is the exposure or treatment?
2. What is the comparator?
3. Who is the target population?
4. What is the outcome?
5. What is the time horizon?
6. What is the causal estimand?
7. What is time zero?
8. What would the counterfactual outcome be?
9. How is treatment assigned?
10. Which variables cause both treatment and outcome?
11. Which variables are mediators?
12. Which variables are colliders?
13. Which variables affect selection?
14. Is randomisation possible?
15. If not, what creates quasi-random variation?
16. Is positivity plausible?
17. Is treatment well-defined?
18. Is there interference?
19. How is missingness generated?
20. Are measurement errors differential?
21. What effect scale matters?
22. Is heterogeneity expected?
23. What mechanism is proposed?
24. What negative control could detect bias?
25. What sensitivity analysis tests hidden confounding?
26. Can the effect transport?
27. What happens at scale?
28. What unintended outcomes matter?
29. What evidence would overturn the conclusion?
30. What is the next strongest causal test?

---

294. Causal Ladder for Students

Level 1 — Notice association

“X and Y change together.”

Level 2 — Ask direction

“Could X cause Y, Y cause X, or Z cause both?”

Level 3 — Establish time

“Did X happen before Y?”

Level 4 — Build fair comparison

“What stays the same while X changes?”

Level 5 — Consider alternatives

“What else could explain the outcome?”

Level 6 — Trace mechanism

“What process connects X to Y?”

Level 7 — Use counterfactual

“What would have happened if X had not changed?”

Level 8 — Recognise confounding

“Which pre-existing factors influence both treatment and outcome?”

Level 9 — Use causal design

“Which randomised or quasi-experimental comparison identifies the effect?”

Level 10 — Transport and intervene

“Will changing X produce the same effect here, at this scale, in this population?”

This ladder can develop from Primary Science fair-test reasoning into advanced causal inference.

---

295. Causal Claim Strength Ladder

Tier A — Descriptive

“X and Y were observed together.”

Tier B — Associational

“X is statistically associated with Y after measured adjustment.”

Tier C — Temporally ordered

“X precedes Y and remains associated after adjustment.”

Tier D — Quasi-experimental

“Exogenous variation in X changes Y under defended assumptions.”

Tier E — Randomised causal

“Assignment to X changes Y relative to control.”

Tier F — Mechanistically convergent

“Randomised or quasi-experimental effect is supported by the proposed pathway and independent evidence.”

Tier G — Transported and replicated

“Causal effect replicates across populations, designs and mechanisms with known heterogeneity.”

This is not a universal ranking system for all science. It is a teaching ladder showing how additional design and evidence can strengthen causal claims.

---

296. Causal Versioning

Causal claims should be versioned when the intervention, domain or mechanism changes.

Example:

Version 1: Treatment A reduces Outcome Y.

Version 2: Treatment A reduces Y only in Population P.

Version 3: Treatment A reduces Y in P through Mediator M, with little effect in subgroup Z.

Each version is more precise.

Preserve the old versions so the evidential history remains visible.

297. Causal retirement rule

A causal claim should be reconsidered or retired when:

• multiple well-designed studies fail;
• hidden bias explains earlier results;
• a rival causal model predicts evidence better;
• the proposed intervention has no effect in direct tests;
• the mechanism repeatedly fails;
• the effect disappears in better measurements.

Retirement is scientific progress.

---

298. Causal Stewardship Standard

A mature causal claim should be stewarded over time.

Stewardship includes:

• updating effect estimates;
• monitoring new populations;
• checking mechanism;
• tracking treatment versions;
• reassessing harms;
• monitoring spillovers;
• testing scale-up.

A cause–effect claim can change as systems change.

299. Causal reasoning standard

Future updates to this guide should deepen the central question:

how science decides whether changing X would change Y.

They should strengthen:

• counterfactual reasoning;
• confounding control;
• causal design;
• mechanism;
• transport;
• interference;
• causal attribution.

Experiments, evidence, explanation, hypotheses, theories and prediction remain related but distinct topics.

300. The causal contract

A serious causal claim makes a contract with the reader:

I will tell you what intervention I mean, what comparison defines the effect, what assumptions make that comparison credible, which alternative paths could create the association, what mechanism supports the effect, which population the claim applies to, and what new evidence would make me revise it.

That contract is stronger than saying two variables are correlated.

It turns observation into an actionable scientific claim.

301. Final close

Causation is the part of science that lets knowledge become intervention.

Without causal reasoning, we can describe the world and predict parts of it.

With causal reasoning, we can ask whether changing one part will change another.

That is why causation matters in medicine, education, engineering, environment and policy.

But causal language must be earned.

The full discipline is:

observe the association → define the intervention → imagine the counterfactual → map confounding and selection → build a credible comparison → estimate the effect → test the mechanism → examine heterogeneity → transport cautiously → monitor what happens after action.

Science becomes causal not when it uses the word “because”, but when the comparison between possible worlds becomes credible enough that changing X really means something about Y.

---

302. Causal Verification Appendix

A causal estimate becomes scientifically useful when it can be verified from the design outward. Verification should not begin with the final coefficient. It should begin with the causal question and reconstruct every link that made the coefficient interpretable.

Verification 1 — Reconstruct the intervention

Write exactly what was changed. “Treatment” is too vague if different participants received different versions. Record dose, duration, timing, delivery and adherence rules.

Verification 2 — Reconstruct the comparator

State what treatment was compared against. No treatment, placebo, standard care and alternative treatment answer different causal questions.

Verification 3 — Reconstruct time zero

Eligibility, treatment assignment and start of follow-up should align. If they do not, causal estimates can inherit immortal-time and selection biases.

Verification 4 — Reconstruct the population

Identify who was eligible, who entered the study, who remained, and who was analysed. Each filter can alter the target population.

Verification 5 — Reconstruct the causal graph

List the pre-treatment common causes, mediators, selection variables and potential colliders. The adjustment set should follow this structure rather than software defaults.

Verification 6 — Reconstruct the identification argument

Explain why the observed comparison identifies the counterfactual contrast. In randomized trials, cite the assignment mechanism. In observational studies, state the no-unmeasured-confounding or quasi-experimental assumptions.

Verification 7 — Reconstruct the estimator

Document matching, weighting, regression, g-computation or another estimator. Estimation should be reproducible after identification has been justified.

Verification 8 — Reconstruct the effect scale

State whether the result is a risk difference, risk ratio, odds ratio, mean difference, hazard ratio or another estimand. Do not compare effects across scales casually.

Verification 9 — Reconstruct uncertainty

Include sampling uncertainty, model sensitivity and key identification assumptions. A narrow confidence interval does not represent uncertainty from unmeasured confounding automatically.

Verification 10 — Reconstruct sensitivity

Show how the conclusion changes under reasonable alternative specifications, missing-data assumptions, confounding strengths or treatment definitions.

303. Causal Replication Appendix

Replication of causation should ask more than whether the same p-value reappears.

A causal replication should compare:

• the same intervention contrast;
• the same outcome definition;
• the same follow-up period;
• the same or intentionally different population;
• the same mechanism;
• the same identification strategy.

If the effect differs, investigate effect modification before calling the replication a failure.

Replication by design diversity

The causal case becomes especially strong when a randomized trial, natural experiment, longitudinal observational study and mechanistic experiment converge on the same conclusion. Each design answers with different vulnerabilities.

Replication by population diversity

Testing different populations maps external validity. A changing effect is scientifically informative if it tracks identifiable moderators rather than random noise.

Replication by intervention version

Two implementations of “the same” programme may differ enough to produce different effects. Implementation fidelity belongs inside causal interpretation.

304. Causal Transport Appendix

Transporting causal knowledge is one of the most important steps between research and real-world use.

Ask:

• Which variables modify the effect?
• How do their distributions differ between study and target?
• Is treatment delivered the same way?
• Are baseline risks different?
• Are competing pathways present?

If important effect modifiers are measured, researchers can sometimes standardise or reweight the study estimate to the target population.

If important modifiers are unmeasured, transport uncertainty should remain explicit.

305. Scale-Up Appendix

A treatment that works for 100 people may behave differently when applied to one million.

Scale-up can introduce:

• resource constraints;
• provider learning;
• spillovers;
• market responses;
• behavioural adaptation;
• general-equilibrium effects.

The large-scale causal effect is therefore not always the small-trial effect multiplied by population size.

Capacity constraints

A tutoring intervention can work in a small study with highly trained instructors but weaken when scaled because tutor quality changes.

Spillovers

A vaccination programme can become more effective at scale because transmission falls.

Behavioural adaptation

A road policy can change route choice and land use over time.

Scale-up is itself a new causal experiment.

306. Causation and Decision Thresholds

Causal estimates do not make decisions automatically.

Decision-makers must combine:

• effect size;
• uncertainty;
• cost;
• harm;
• equity;
• feasibility.

A small causal effect can be worthwhile if the intervention is cheap and the population is large.

A large causal effect can be impractical if harms or costs are unacceptable.

307. Benefit–harm causation

Interventions can cause both good and bad outcomes.

A complete causal assessment should estimate:

• benefit;
• adverse effects;
• dropout;
• spillovers;
• long-term consequences.

Do not define success using only the outcome the intervention was designed to improve.

308. Causal evidence and mechanism should meet before policy

Policy confidence becomes stronger when:

• causal effect is demonstrated;
• mechanism is credible;
• effect transports;
• implementation is realistic;
• unintended effects are monitored.

Mechanism helps anticipate where the intervention may fail after scale-up.

309. Causal confidence levels

Preliminary

Association plus temporal plausibility, with important confounding unresolved.

Suggestive

Multiple adjusted studies or quasi-experimental signals point in the same direction.

Strong

Credible intervention or natural-experiment evidence with supportive mechanism and replication.

Very strong within domain

Convergent randomized, observational and mechanistic evidence across populations with known effect heterogeneity and boundaries.

These labels are descriptive, not universal formal categories. Their purpose is to keep causal language proportional to evidence.

310. Final Causal Verification Checklist

• The intervention is clearly defined.
• The comparator is explicit.
• The target population is named.
• Time zero is aligned.
• The outcome and horizon are defined.
• The causal estimand is stated.
• The assignment mechanism is understood.
• The DAG is explicit or causal assumptions are otherwise documented.
• Confounders are measured and justified.
• Mediators and colliders are not adjusted blindly.
• Selection and missingness are assessed.
• Interference is considered.
• Treatment versions are considered.
• Effect heterogeneity is assessed where scientifically motivated.
• Sensitivity analyses address hidden bias.
• Mechanism evidence is distinguished from effect evidence.
• Absolute and relative effects are interpreted appropriately.
• Transport assumptions are visible.
• Scale-up effects are not assumed automatically.
• The causal claim remains revisable.

311. Final causal close

Scientific causation is the discipline that turns description into intervention.

It asks more of evidence than correlation, more of design than prediction, and more of interpretation than statistical significance.

The full causal argument is strongest when it can answer:

What was changed? Compared with what? Why is the comparison credible? Which alternative paths were blocked? What mechanism connects cause to effect? For whom does the effect apply? What happens when the intervention scales? What evidence would make us change our mind?

When those questions are answered clearly, cause-and-effect language becomes an accountable scientific claim rather than a rhetorical shortcut.

---

312. Causal Assumption Matrix

Every causal estimate rests on assumptions. Making them explicit is one of the clearest differences between serious causal inference and ordinary statistical association.

Assumption	Main question	What threatens it?
Temporality	Did the cause precede the effect?	cross-sectional measurement, reverse causation
Exchangeability	Are treatment groups comparable in potential outcomes?	confounding, selection
Consistency	Does observed treatment correspond to the intervention being defined?	multiple treatment versions
Positivity	Could relevant units receive either treatment?	structural treatment rules
Measurement validity	Are exposure and outcome measured correctly?	misclassification, instrument bias
No harmful selection	Does sample inclusion distort treatment–outcome relations?	dropout, collider selection
Model adequacy	Does the estimator represent the data well enough?	misspecification, extrapolation
Transportability	Do study effects apply to target population?	effect modification, treatment differences

313. Which assumptions can data test directly?

Some assumptions leave observable implications.

Examples:

• pre-treatment balance can be checked in randomized trials;
• parallel pre-trends can be inspected in difference-in-differences;
• instrument strength can be measured;
• measurement reliability can be audited;
• positivity violations can be seen in treatment probabilities.

Other assumptions are partly or largely untestable from the same data.

Examples include:

• no unmeasured confounding;
• valid exclusion restriction for an instrument;
• absence of unknown selection pathways.

Untestable does not mean arbitrary. It means the assumption must be defended using subject knowledge, design, external evidence and sensitivity analysis.

314. Causal Evidence Matrix

Evidence type	What it contributes	Main limitation
Randomized trial	strong treatment contrast	implementation/generalisation
Cohort study	temporality, long-term effects	confounding
Case-control study	efficient rare-outcome study	selection/recall bias
Natural experiment	quasi-random exposure variation	assignment assumptions
Regression discontinuity	local causal effect near threshold	locality/manipulation
Difference-in-differences	policy effect from trend contrast	parallel-trends assumption
Instrumental variable	causal estimate under hidden confounding	strong validity assumptions
Mechanistic experiment	pathway evidence	may not show population effect
Negative control	detects hidden bias	requires valid control choice
Triangulation	reduces shared-bias risk	methods may not target identical estimands

315. Causal claims should match the strongest supported level

Scientific writing becomes clearer when causal wording matches design.

Examples:

Descriptive: “The outcome was higher in the exposed group.”

Associational: “Exposure was associated with higher outcome after adjustment.”

Quasi-causal: “The policy change was followed by an outcome change relative to the comparison trend under the parallel-trends assumption.”

Causal experimental: “Assignment to treatment reduced the outcome relative to control.”

Mechanistically causal: “Treatment reduced the outcome through Pathway M, supported by mediation and pathway-interruption evidence.”

The language should become stronger only as the design earns it.

316. The Causal Stop Rule

Researchers need to know when evidence is strong enough to act, even if uncertainty remains.

A practical stop rule can ask:

• is the causal effect replicated?
• is the magnitude decision-relevant?
• are major biases unlikely to reverse the conclusion?
• does mechanism support transport?
• are harms understood?
• does waiting for perfect certainty create greater harm?

This is not a statistical formula.

It is an evidence-to-decision boundary.

317. The Causal Reopen Rule

A causal claim should be reopened when:

• new populations show different effects;
• better measurement changes the association;
• a hidden confounder is discovered;
• the mechanism fails replication;
• scale-up changes behaviour;
• a rival design produces conflicting evidence.

Causal science remains alive after publication.

318. The final causal standard

The strongest causal claim is not the one stated most confidently.

It is the one whose counterfactual is clearest, whose design makes the comparison credible, whose assumptions are visible, whose mechanism is testable, whose effects replicate, and whose boundaries are known.

That is the standard this guide should preserve.

---

319. Causal Communication Standard

Causal inference can be technically rigorous and still be communicated badly. Public-facing causal claims should preserve the distinctions that made the analysis valid.

Say what was compared

Instead of:

“The treatment works.”

Prefer:

“Assignment to Treatment A reduced Outcome Y compared with Standard Care over six months in the study population.”

The comparator and time horizon belong in the causal statement.

Say who the claim applies to

A causal effect estimated in one population should not be silently universalised. State the study population and identify important differences in the target population.

Say whether the claim is experimental or observational

Readers should know whether treatment was randomised, naturally assigned or observed. This changes how strongly unmeasured confounding must be considered.

Say what assumptions matter most

Do not hide the key causal assumption behind technical terminology. Examples:

• groups would have followed similar trends without the policy;
• the instrument affects the outcome only through treatment;
• no important unmeasured confounder remains after adjustment.

Say how large the effect is

“Statistically significant” is not a causal effect size.

Where relevant, report:

• absolute risk difference;
• relative risk;
• number affected per 1,000 people;
• mean outcome difference.

Say how uncertain the estimate is

Include intervals and sensitivity analyses. A precise sampling interval does not eliminate uncertainty from causal assumptions.

Say what remains alternative

If unmeasured confounding, selection or a competing mechanism remains plausible, state that boundary openly.

Say whether the mechanism is established

An intervention can have a causal effect before its pathway is fully known. Do not turn “X affects Y” into “X affects Y through M” unless M has evidence.

Say whether the effect transported

If the causal effect has been replicated in different populations or settings, say so. If not, avoid implying universal transfer.

Say what would make the conclusion change

A strong causal statement remains revisable. Identify the kind of evidence that would weaken or overturn it.

320. Causation in one sentence

The entire owner can be compressed into one operational question:

If we deliberately changed X from one defined state to another, for this population and under these conditions, how would Y differ from what would otherwise have happened?

Everything else—randomisation, controls, causal diagrams, confounding, mediation, natural experiments, transport and mechanism—exists to make that comparison more credible.

321. Final integrity line

A causal claim should never be stronger than its counterfactual design. The honest scientific habit is to distinguish what was observed, what was predicted, what was intervened upon, what assumptions connected the comparison, and what evidence could still show that the causal story is incomplete.

---

322. Causal Boundary and Reassessment Standard

A causal claim should always carry a boundary map. The effect established in one experiment or population is not automatically the effect everywhere. Boundaries can arise from dose, timing, biology, environment, implementation, baseline risk or system scale.

For every mature causal claim, record:

• where the effect has been observed;
• where it has been replicated;
• where the effect is weaker or stronger;
• where the causal mechanism changes;
• where the intervention has not been tested;
• which populations remain uncertain.

This map prevents two opposite mistakes. The first is overgeneralisation: assuming that one successful causal estimate is universal. The second is undergeneralisation: treating every context as unrelated and failing to identify a transportable mechanism.

Reassessment after new evidence

A causal conclusion should be revisited when new evidence changes one of the structural assumptions. Examples include a newly discovered confounder, a revised measurement system, a subgroup with an opposite effect, or a scale-up that introduces spillovers.

The reassessment sequence is:

1. preserve the original causal estimate and assumptions;
2. identify which assumption the new evidence challenges;
3. re-estimate the effect under the revised structure;
4. compare whether the causal conclusion, magnitude or domain changed;
5. publish the revised boundary rather than hiding the discrepancy.

Causal science becomes more trustworthy when its revisions remain visible.

Causal continuity

This causation guide should continue to answer the intervention question—what changes what—while related guides treat their separate scientific questions. Evidence determines confidence, experiments build tests, explanations describe mechanisms, hypotheses frame candidate claims, prediction forecasts unseen outcomes, and theories organise the larger framework.

The final public standard is therefore:

Causal confidence should rise only when the intervention, comparison, assumptions, mechanism, replication and transport all remain visible enough for another scientist to challenge them.

323. Final causal continuity note

This guide should remain useful even as methods evolve. New estimators, machine-learning tools and quasi-experimental designs may change how causal effects are calculated, but the durable scientific obligations remain the same: define the intervention, identify the counterfactual, justify the comparison, expose assumptions, test the mechanism, measure heterogeneity, verify transport, preserve failures and revise openly when stronger evidence changes the causal story.

Integrity: Causal claims remain scientific only while stronger evidence can still revise their stated scope, magnitude, mechanism and domain.