Module 3 — Correlation vs Causation

Why relationships in healthcare data are rarely straightforward

When two things move together, does one cause the other?

Healthcare analytics frequently identifies relationships between variables.

For example:

higher deprivation and worse health outcomes
frailty and increased admissions
obesity and diabetes
delayed discharge and longer Length of Stay
smoking and stroke incidence

When two variables appear related, this is known as a correlation.

However, correlation alone does not prove that one variable directly causes the other.

This distinction matters enormously in healthcare analytics because healthcare systems are complex, interconnected and influenced by many overlapping factors.

What Is Correlation?

Imagine we observe:

Area	Fast-food density	Obesity
Area A	Low	Low
Area B	High	High

We might conclude:

fast-food outlets directly cause obesity

But what else may matter?

income
access to healthy food
physical activity
deprivation
housing
education

The relationship may be real — but the pathway behind it is often more complex than it first appears.

Correlation describes a relationship between two variables.

For example:

Observation	Relationship
Areas with higher deprivation often have higher stroke mortality	Positive correlation
Increased physical activity is associated with lower cardiovascular risk	Negative correlation

Correlation helps us:

identify patterns
detect variation
generate hypotheses
prioritise investigation

But correlation alone cannot tell us:

why the relationship exists
whether the relationship is direct
whether other factors are involved

Correlation Does Not Automatically Mean Causation

Suppose we observe:

Area	Stroke Mortality
Most deprived areas	Higher
Least deprived areas	Lower

It may be tempting to conclude:

“Deprivation directly causes higher stroke mortality.”

However, healthcare outcomes are rarely driven by a single factor.

Deprivation may instead be linked to:

higher smoking prevalence
hypertension
obesity
diabetes
delayed healthcare access
poorer housing
reduced access to prevention
increased multimorbidity
environmental factors

In reality, deprivation often acts as a marker for a wider set of underlying risks and system pressures.

Example — Stroke Outcomes and Inequalities

Stroke outcomes often vary significantly between populations.

For example, more deprived populations may experience:

higher stroke incidence
earlier onset of stroke
increased mortality
longer rehabilitation needs
poorer long-term outcomes

However, these differences are usually influenced by multiple interconnected factors including:

prevalence of hypertension
smoking rates
access to primary care
anticoagulation gaps in atrial fibrillation
housing and social conditions
delayed presentation to services
rehabilitation access
transport and geography

This means:

the observed inequality is real, but the pathway causing it is complex.

The example below illustrates how healthcare inequalities often emerge through multiple interconnected risk factors rather than a single direct cause-and-effect relationship.

Poor Interpretation vs Better Interpretation

Observed:

higher deprivation → higher stroke mortality

Poor interpretation:

“Deprivation causes stroke.”

Better interpretation:

“Deprivation is associated with a set of behavioural, environmental, biological and access-related risks that may contribute to stroke outcomes.”

Confounding Variables

A confounding variable is a factor that influences both variables being studied.

For example:

Deprivation
      ↓
Smoking prevalence
      ↓
Stroke risk

If smoking prevalence is not considered, we may overestimate the direct effect of deprivation alone.

Healthcare datasets often contain many confounding variables simultaneously.

This is one reason why interpreting healthcare analytics requires caution.

Why Adjustment Matters

Healthcare analytics often uses adjustment methods to improve fairness and interpretation.

Examples include:

age standardisation
case-mix adjustment
regression modelling
risk adjustment

Adjustment helps us ask:

“What would outcomes look like if populations were more comparable?”

For example, after adjusting for:

age
smoking
hypertension
diabetes
frailty

…the gap in stroke mortality between populations may reduce substantially.

This does not mean inequalities disappear.

It means:

part of the observed relationship was explained by other linked factors.

Adjustment improves understanding — it does not remove complexity.

Looking Deeper

If interested, see:

Appendix — How Adjustment Works

for a more detailed explanation of adjustment approaches used in healthcare analytics.

Correlation Can Still Be Extremely Useful

Importantly:

correlation is not “bad”.

Correlation is often the starting point for:

identifying inequalities
detecting unwarranted variation
prioritising interventions
operational investigation
pathway redesign

Without correlation analysis:

many healthcare inequalities would remain hidden.

The key is understanding that:

correlation should start questions, not end them.

Operational Implications

Misinterpreting correlation as causation can lead to:

simplistic conclusions
ineffective interventions
unrealistic targets
inappropriate benchmarking

For example:

If analysis shows:

higher ED attendance in deprived areas

…the operational response should not simply be:

“reduce ED demand.”

Instead, systems may need to examine:

primary care access
prevention services
frailty support
transport barriers
community capacity
housing instability
proactive care pathways

This moves analytics from:

descriptive reporting towards:
operational understanding and intervention.

From Correlation to Operational Action

Effective healthcare analytics often follows this progression:

Step	Purpose
Correlation	Detect a signal
Identify drivers	Explore explanations
Adjustment	Improve interpretation
Cohort segmentation	Identify who is affected
Operational targeting	Focus intervention
Pathway redesign	Improve outcomes

This is where healthcare analytics becomes most valuable — moving from observation to better intervention design.

Why Healthcare Data Is Particularly Complex

Healthcare outcomes rarely arise from a single cause.

Instead:

multiple factors interact
pathways overlap
operational context matters

This is why healthcare analytics requires:

interpretation, context and operational understanding — not statistical outputs alone

Key Takeaways

Correlation describes relationships between variables
Correlation alone does not prove causation
Correlation should start questions, not end them
Healthcare inequalities are usually driven by multiple interacting factors
Confounding variables can distort interpretation
Adjustment helps improve fairness and understanding
Correlation remains extremely valuable for identifying patterns and prioritising investigation
Healthcare analytics should support better questions, not simplistic conclusions

Questions Decision-Makers Should Ask

When reviewing correlated healthcare metrics:

What factors may sit behind this relationship?
Are important confounding variables being missed?
Does this represent causation, association, or both?
Are populations truly comparable?
What operational factors influence these outcomes?
Which drivers are actually modifiable?
What intervention opportunities exist?
What additional analysis is needed before acting?