Association, Causation, Observational Studies, and Randomized Experiments

When two variables occur together, statisticians say they are associated. An association can be useful and scientifically important, but it does not by itself establish that one variable causes the other.

This distinction is fundamental:\[ \boxed{\text{Association does not necessarily imply causation}} \]

Determining causation requires more than comparing outcomes between groups. Researchers must consider confounding, selection, measurement, time order, alternative explanations, and the design of the study.

Association Between Two Variables

Two variables are associated when the distribution of one differs depending on the value of the other.

Suppose researchers compare a health outcome between:

• people reporting higher red-meat consumption;
• people reporting little or no red-meat consumption.

Let:\[ Y= \begin{cases} 1, & \text{if the health outcome occurs}\\ 0, & \text{otherwise} \end{cases} \]

and let \(X\) represent reported consumption.

An observed association might take the form:\[ P(Y=1\mid X=\text{higher}) > P(Y=1\mid X=\text{lower}). \]

This means the outcome rate is higher in one observed group. It does not yet explain why.

Why Association Is Not Automatically Causation

People who differ in one behavior often differ in many other ways.

Groups with different dietary patterns might also differ in:

• age;
• physical activity;
• alcohol use;
• smoking;
• income;
• healthcare access;
• overall diet;
• occupation;
• location;
• existing health conditions.

These additional variables may influence the outcome and may also be related to the exposure under study.

Therefore, the observed difference could arise from:

1. a causal effect of the exposure;
2. another variable affecting both exposure and outcome;
3. reverse causation;
4. selection bias;
5. measurement error;
6. random variation;
7. some combination of these mechanisms.

An association is evidence of a relationship, but its causal interpretation requires further justification.

Confounding Variables

A confounder is a variable associated with both the exposure and the outcome that can create or distort their observed relationship.

Consider:

• \(X\): dietary exposure;
• \(Y\): health outcome;
• \(Z\): smoking behavior.

A possible causal structure is:

        Smoking
        /     \
       v       v
Dietary pattern   Health outcome

In causal notation:\[ Z\rightarrow X \]

and\[ Z\rightarrow Y. \]

If smoking differs between dietary groups and affects the outcome, comparing the groups without adequately addressing smoking may incorrectly attribute some of its effect to diet.

A confounder is sometimes informally called a lurking variable, although causal analysis requires a more precise definition than simply “a variable that was not measured.”

Confounding Can Change the Result

Suppose an observational dataset reports:

The crude difference is:\[ 12\%-8\%=4 \]

percentage points.

Now suppose the higher-exposure group contains a larger proportion of older participants, and age strongly affects the outcome. After comparing participants within similar age groups, the difference might become smaller.

The unadjusted comparison mixes together:

• differences associated with the exposure;
• differences associated with age;
• differences associated with other variables.

Statistical adjustment attempts to separate these contributions, but it succeeds only under appropriate assumptions and adequate measurement.

Observational Studies

In an observational study, researchers observe exposures and outcomes without assigning the exposure under investigation.

Examples include:

• comparing health outcomes across dietary patterns;
• studying smoking histories and disease;
• examining air pollution and respiratory health;
• relating exercise habits to cardiovascular outcomes;
• comparing educational experiences and income.

The defining property is:\[ \text{researchers observe rather than assign the exposure}. \]

Observational studies may be prospective, following participants forward in time, or retrospective, using existing histories and records.

What Observational Studies Can Establish

Observational studies can provide evidence about:

• associations;
• prevalence;
• incidence;
• risk factors;
• patterns in real populations;
• rare or long-term outcomes;
• exposures that cannot ethically be assigned.

They are indispensable in public health, economics, social science, and many other fields.

However, causal interpretation requires strong assumptions because exposure groups may differ before observation begins.

Researchers may address confounding using:

• stratification;
• matching;
• regression adjustment;
• standardization;
• weighting;
• natural experiments;
• instrumental variables;
• interrupted time-series designs;
• sensitivity analysis.

These methods can support causal conclusions under explicit assumptions, but they do not turn observational data automatically into randomized data.

Randomized experiments are powerful tools for causal inference, but carefully designed observational studies can also contribute causal evidence when experiments are infeasible or unethical.

Reverse Causation

Sometimes the apparent direction of cause and effect is reversed.

Suppose an observational study finds:\[ X\leftrightarrow Y. \]

The intended interpretation may be:\[ X\rightarrow Y. \]

But it is possible that:\[ Y\rightarrow X. \]

For example, people with early symptoms may change their diet, exercise, medication, or employment before receiving a formal diagnosis. The outcome process may therefore influence the observed exposure.

Establishing that the exposure occurred before the outcome helps, but temporal order alone does not eliminate confounding.

Randomized Experiments

In a randomized experiment, researchers assign eligible participants to study conditions using a random mechanism.

A basic design has:

• a treatment group;
• a control or comparison group;
• random assignment;
• predefined outcomes;
• consistent follow-up.

Let:\[ T_i= \begin{cases} 1, & \text{if participant }i\text{ receives treatment}\\ 0, & \text{if participant }i\text{ receives control} \end{cases} \]

Random assignment determines \(T_i\), often using a computer-generated random sequence.

The outcome \(Y_i\) is then compared between the groups.

Treatment and Control Groups

Treatment is a technical term referring to the condition being evaluated. It does not necessarily mean medical care.

Treatments can include:

• a medication;
• an educational method;
• a notification design;
• a pricing policy;
• an exercise program;
• a public-information campaign;
• a software interface.

The control group provides a reference for what might have happened without the experimental treatment.

Possible controls include:

• no intervention;
• usual care;
• a placebo;
• an existing treatment;
• a different intervention;
• a wait-list condition.

The appropriate control depends on the scientific question and ethical requirements.

Why Random Assignment Matters

Participants differ in countless ways, including variables that researchers cannot measure.

Random assignment makes treatment status independent of participant characteristics in expectation:\[ T\perp \text{baseline characteristics} \]

under a correctly implemented design.

This does not guarantee that every realized experiment will have perfectly identical groups. Chance imbalances can still occur, especially in small studies.

Rather, randomization ensures that systematic assignment is not based on:

• disease severity;
• age;
• motivation;
• income;
• clinician preference;
• participant choice;
• predicted outcome.

As sample size increases, substantial baseline imbalances generally become less likely, although the actual balance should still be examined.

Potential Outcomes and the Causal Effect

For each participant \(i\), imagine two potential outcomes:\[ Y_i(1)=\text{outcome if treated} \]\[ Y_i(0)=\text{outcome if not treated}. \]

The individual causal effect is:\[ Y_i(1)-Y_i(0). \]

The fundamental difficulty is that only one outcome can be observed for the same participant:\[ Y_i= T_iY_i(1)+(1-T_i)Y_i(0). \]

If the participant receives treatment, \(Y_i(1)\) is observed but \(Y_i(0)\) is not. If the participant receives control, the reverse is true.

Random assignment creates comparable groups so that the average outcome in one group can estimate the missing counterfactual outcome for the other group.

The average treatment effect is:\[ \operatorname{ATE} = E[Y(1)-Y(0)]. \]

A randomized comparison estimates it with:\[ \widehat{\operatorname{ATE}} = \bar{Y}_{T=1} – \bar{Y}_{T=0}. \]

Random Assignment Versus Random Sampling

These two forms of randomization have different purposes.

Random sampling

Random sampling selects participants from a population.

It supports generalization:\[ \text{sample}\rightarrow\text{population}. \]

Random assignment

Random assignment allocates participants to study conditions.

It supports causal comparison within the study:\[ \text{treatment}\rightarrow\text{outcome}. \]

An experiment can have strong internal causal validity even if its participants are not a random sample of the broader population. Its results may nevertheless generalize poorly to other populations.

Allocation Concealment

The random sequence should not be predictable before a participant is enrolled.

Allocation concealment prevents recruiters from knowing the upcoming condition and consciously or unconsciously influencing who enters the study next.

For example, using a visible alternating sequence such as

Treatment, Control, Treatment, Control

is not adequate randomization because the next assignment is predictable.

Allocation concealment protects the assignment process before and during enrollment. It is distinct from blinding, which protects the study after assignment.

Placebo Controls

A placebo resembles the experimental treatment but lacks its specific active component.

For a medication study, a placebo might resemble the treatment pill in:

• appearance;
• taste;
• packaging;
• administration schedule.

A placebo-controlled design attempts to separate:

• the specific effect of the active treatment;
• expectations about receiving care;
• natural changes over time;
• attention from researchers;
• reporting differences.

The control group undergoes an experience similar to that of the treatment group, except for the active component being evaluated.

The Placebo Effect

The phrase placebo effect is often used broadly, but several mechanisms can contribute to improvement observed after placebo administration:

• expectations;
• learning and conditioning;
• changes in symptom perception;
• interaction with clinicians;
• regression toward the mean;
• natural recovery;
• changes in reporting behavior.

Not every improvement in a placebo group is caused by the placebo itself.

To isolate a placebo response from natural history and other effects, a study may need both:

• a placebo group;
• a no-treatment or usual-care group.

The placebo response can be meaningful for subjective symptoms such as perceived pain or nausea, but it should not be assumed to affect every biological outcome in the same way.

Placebos Are Not Always Appropriate

A placebo may be unethical when withholding an effective established treatment would expose participants to serious harm.

In such cases, the comparison may be:\[ \text{new treatment} \quad\text{versus}\quad \text{standard treatment}. \]

Other possibilities include:

• adding the new treatment to usual care;
• comparing two accepted treatments;
• using an active placebo that mimics side effects;
• using a wait-list control when ethically acceptable.

Scientific rigor must be balanced with participant safety and informed consent.

Blinding

Blinding means withholding treatment-assignment information from people whose expectations or decisions could affect the study.

Possible groups to blind include:

• participants;
• clinicians;
• outcome evaluators;
• laboratory personnel;
• data analysts.

Terminology such as single-blind and double-blind is used inconsistently. It is clearer to state exactly who was blinded.

For example:

Participants and outcome evaluators were blinded to treatment allocation.

This is more informative than writing only “double-blind.”

Why Blinding Matters

Knowing the assigned condition can affect:

Participant behavior

Participants may change:

• symptom reporting;
• adherence;
• use of additional treatments;
• expectations;
• willingness to remain in the study.

Clinician behavior

Clinicians may provide different:

• encouragement;
• follow-up;
• additional care;
• interpretation of symptoms.

Outcome assessment

Evaluators may unconsciously rate ambiguous outcomes more favorably for the treatment group.

Blinding reduces these pathways of bias.

Blinding Is Not Always Possible

Some interventions cannot be concealed.

Examples include:

• surgery versus no surgery;
• psychotherapy;
• exercise programs;
• educational methods;
• workplace policies;
• user-interface changes.

When participant or provider blinding is impossible, researchers can still use:

• blinded outcome evaluators;
• objective outcome measures;
• standardized protocols;
• automated measurements;
• blinded statistical analysis;
• predefined decision rules.

The goal is to block as many bias pathways as the design permits.

Blocking and Stratified Randomization

Pure random assignment can produce chance imbalance in small studies.

Researchers may first group participants according to an important baseline variable and randomize within each group.

For example, participants might be divided by:

• age category;
• disease severity;
• study center;
• sex;
• prior treatment;
• risk level.

This is called blocking or stratified randomization.

Within block \(h\), participants are assigned independently to treatment and control. This helps maintain balance on variables considered especially important.

It serves a role similar to stratified sampling, but the objective differs:

• stratified sampling improves population estimation;
• stratified randomization improves treatment-group comparability.

Paired and Crossover Designs

Matched-pair design

Participants are paired based on similar characteristics. Within each pair, one receives treatment and the other receives control.

Alternatively, the same participant may be measured before and after treatment.

Crossover design

Participants receive multiple conditions in randomized order:\[ A\rightarrow B \]

or\[ B\rightarrow A. \]

Each participant can serve as their own control.

Crossover designs are useful only when:

• treatment effects are temporary;
• sufficient washout is possible;
• order effects can be addressed;
• the condition is stable enough for repeated measurement.

They are unsuitable when the treatment creates a permanent change.

Adherence and Treatment Switching

Random assignment does not guarantee that participants follow their assigned condition.

Possible complications include:

• participants not taking the treatment;
• control participants obtaining a similar treatment elsewhere;
• participants switching groups;
• participants leaving the study;
• missing outcome measurements.

A common primary analysis follows the intention-to-treat principle:

Analyze participants according to their randomized groups, regardless of full adherence.

This preserves the benefits of random assignment and estimates the effect of being assigned the treatment under the study conditions.

An analysis based only on participants who followed the protocol may reintroduce confounding because adherence itself can be related to prognosis.

Internal and External Validity

Internal validity

Internal validity concerns whether the observed difference can credibly be attributed to the assigned treatment within the study.

Threats include:

• failed randomization;
• differential dropout;
• unblinded assessment;
• treatment contamination;
• missing outcomes;
• inconsistent measurement.

External validity

External validity concerns whether the results apply to other people, places, treatments, or periods.

Threats include:

• narrow eligibility rules;
• volunteer participation;
• unusual clinical settings;
• short follow-up;
• different treatment delivery;
• underrepresentation of important groups.

A study may have high internal validity but limited external validity.

Experiments Do Not Automatically Prove Every Causal Claim

Randomized experiments are powerful, but interpretation still depends on design and implementation.

Problems can arise from:

• noncompliance;
• dropout;
• missing data;
• inadequate sample size;
• multiple unplanned comparisons;
• faulty measurement;
• compromised blinding;
• treatment spillovers;
• selective reporting;
• short follow-up;
• chance imbalances.

Randomization provides a basis for causal inference. It does not compensate for every other design failure.

When Experiments Are Impossible or Unethical

Researchers cannot randomly assign many exposures.

Examples include:

• smoking;
• harmful pollution;
• traumatic events;
• poverty;
• long-term hazardous diets;
• genetic characteristics.

In such cases, causal evidence may come from a combination of:

• prospective observational studies;
• natural experiments;
• policy changes;
• mechanistic evidence;
• dose–response patterns;
• repeated findings;
• negative controls;
• sensitivity analyses;
• triangulation across study designs.

Causal conclusions are often strongest when multiple methods with different weaknesses point toward the same result.

A Comparison of Study Designs

Neither design is universally superior. The appropriate design depends on the scientific question, feasibility, ethics, and required conclusion.

Evaluating a Causal Claim

When a report says that one factor causes another, ask:

1. Was the study observational or experimental?
2. How were participants selected?
3. Was treatment randomly assigned?
4. What was the control condition?
5. Were allocation and outcome assessment concealed?
6. Which variables could confound the relationship?
7. Were exposure and outcome measured accurately?
8. Did the exposure occur before the outcome?
9. Was dropout different between groups?
10. Were results consistent across groups and analyses?
11. How large was the estimated effect?
12. How uncertain was the estimate?
13. Does the conclusion apply to the intended population?
14. Are alternative explanations plausible?

The study headline should never substitute for examining its design.

Common Misconceptions

“Association proves causation when the sample is large”

A large sample reduces random error but does not remove confounding or systematic bias.

“Adjusting for several variables removes all confounding”

Adjustment addresses measured variables included appropriately in the model. Unmeasured or poorly measured confounders may remain.

“Randomization makes the groups identical”

It makes them comparable in expectation. Chance differences can still occur.

“Every control group should receive a placebo”

An active treatment, usual care, or another comparison may be scientifically and ethically more appropriate.

“Double-blind always has one precise meaning”

The term is used inconsistently. Reports should identify exactly who was blinded.

“Every improvement in a placebo group is a placebo effect”

Natural recovery, regression toward the mean, additional care, and measurement variation can also contribute.

“Randomized findings apply to everyone”

Causal validity within the study does not guarantee generalizability to every population.

Key Takeaway

Association describes a relationship between variables, but causation asks what would happen under an intervention. Observational studies can identify important associations, yet confounding, reverse causation, selection, and measurement problems complicate causal interpretation. Random assignment creates comparable treatment groups in expectation and provides a strong foundation for causal inference. Appropriate controls, allocation concealment, blinding, ethical safeguards, adherence analysis, and careful measurement are still essential. Experiments strengthen causal conclusions, but no study design should be interpreted without examining how it was conducted.