Correlation, Confounding & Causal Inference

When two things move together, several explanations remain possible.

A causes B.

B causes A.

A third factor influences both.

Selection, measurement or chance creates the pattern.

Causal inference is the set of methods used to distinguish among those possibilities.

Correlation is a description, not a mechanism

An association can be important even when it is not causal.

It may identify a useful predictor, a population at higher risk or a pattern worth investigating.

The error occurs when the association is silently upgraded into a causal statement.

That upgrade requires additional assumptions and evidence.

Confounding creates alternative explanations

A confounder is a factor related to both the exposure and the outcome that can produce or distort an observed association.

For example, if coffee drinkers appear to have a higher rate of an outcome, smoking, age, occupation or another variable may partly explain the difference.

Statistical adjustment can help only for variables that are measured and modeled appropriately.

It cannot automatically remove unknown or poorly measured confounding.

Time order matters

A cause must precede its effect.

Cross-sectional data measured at one time point often make direction difficult to establish.

Longitudinal designs can clarify temporal sequence, but timing alone is not enough.

A factor can occur first without causing what follows.

Randomization creates a strong causal contrast

When feasible and well conducted, randomized assignment helps break the link between treatment choice and participant characteristics.

That makes treatment groups more comparable and strengthens causal interpretation.

But randomized trials answer the causal question actually tested — in the studied population, intervention, comparator and timeframe.

They do not automatically generalize beyond it.

Observational data can still support causation

Many causal questions cannot be randomized.

Environmental exposures, long-term behaviors and rare harms often require observational evidence.

Strong observational causal inference can use:

• longitudinal designs;
• natural experiments;
• instrumental variables;
• regression discontinuities;
• negative controls;
• target-trial emulation;
• careful confounder adjustment;
• triangulation across different designs.

No method is magic. Each has assumptions that must be defended.

DAGs make assumptions visible

A directed acyclic graph is one way to represent hypothesized causal relationships before analysis.

Its value is not that the diagram proves causation.

It forces the analyst to state assumptions about which variables cause which others and helps identify:

• variables that should be adjusted for;
• mediators that should not be treated like confounders;
• colliders that can introduce bias if conditioned on;
• paths that remain open.

This makes causal reasoning more inspectable.

Statistical control cannot rescue a bad question

A regression model with many covariates can still estimate the wrong causal quantity.

Before analyzing data, the study must define:

• the exposure;
• the outcome;
• the comparison;
• the population;
• the relevant time period;
• the causal contrast of interest.

Causal inference is a design problem before it is a software problem.

Multiple lines of evidence can strengthen inference

A causal interpretation becomes more credible when different methods with different biases point in the same direction.

For example, randomized trials, natural experiments, mechanistic evidence and dose-response patterns may converge.

Agreement across methods is useful because the same hidden bias is less likely to explain every result.

The Claim by Claim question

When an Investigation encounters an association, we ask:

What alternative process could create the same pattern?

Then we look for designs or evidence that discriminate among those alternatives.

That is the difference between reporting a correlation and making a causal claim.