Does path analysis require latent variables?

Conventional path analysis works with observed variables; broader SEM can include latent measurement models.

Is an indirect path always causal mediation?

No. Causal mediation requires additional assumptions and evidence.

Can different path diagrams fit the same data?

Yes, which limits what fit alone can establish.

AI Glossary / AI and machine learning / Intermediate

Path analysis (statistics)

Definition

Path analysis is a statistical method for estimating a specified network of relationships among observed variables. It extends regression to multiple linked equations and can distinguish direct associations from indirect associations that pass through intermediate variables.

What it means

A study may propose that training relates to confidence, which in turn relates to task performance, while training also has a separate relationship with performance. A path diagram makes that proposed structure visible.

In its conventional form, path analysis uses observed variables rather than a separate latent measurement model. It is closely related to structural equation modeling, which can include broader measurement structures.

Arrows encode assumptions or hypotheses. They do not create causal evidence by being drawn. Different models can sometimes fit the same covariance patterns, and omitted variables, reverse direction and measurement error can undermine an intended causal interpretation.

How it works

The researcher specifies equations and the directions of relationships before estimation. For a simple mediation model, one equation predicts the mediator from the initial predictor, and another predicts the outcome from both.

The fitted coefficients describe relationships under the model. In a linear model without interactions, multiplying the relevant coefficients gives the model's indirect path, and adding direct and indirect paths gives the total relationship under the stated specification.

Assessment considers identification, estimation assumptions, uncertainty and model fit where informative. A model with no degrees of freedom may reproduce the observed covariance structure exactly without providing a meaningful overall fit test.

Three practical examples

A research project examines a proposed relationship among training, confidence and measured task performance.
A survey study separates direct and indirect associations in an explicitly stated conceptual model.
An analyst compares competing path structures and checks whether the available study design can distinguish their interpretations.

Think of it this way

Think of a route map showing both a direct road and a route through an intermediate town. The map makes several possible connections explicit.

The analogy is limited because statistical paths are estimated relationships, not physical pipes carrying a substance. Drawing a route does not establish that influence actually travels along it, and unmeasured connections may change the interpretation.

Worked example

Calculating a simple indirect path

Suppose a linear model estimates the path from X to mediator M as a = 0.5 and the path from M to Y, controlling for X, as b = 0.4. The direct path from X to Y, controlling for M, is c-prime = 0.1.

The model's indirect path is a times b = 0.20. Its total path is 0.1 + 0.20 = 0.30 under this simple linear, no-interaction specification.

These invented coefficients illustrate algebra, not a causal finding. Their uncertainty must be assessed, often through an appropriate bootstrap or another justified method. The product of two estimates can have an uncertainty distribution that is not well described by a naive symmetric approximation.

Even if the indirect path differs statistically from zero, a cross-sectional observational design does not by itself establish the required temporal ordering or rule out confounding. The causal interpretation requires stronger assumptions and evidence.

Illustrative case study

Fictional case: studying a training program

Maple Learning, an invented organization, surveys participants about confidence and records a task score. Analysts fit a path model proposing that training exposure relates to performance partly through confidence.

Because exposure was not randomly assigned and confidence and performance were measured close together, the team reports associations and limitations. It proposes a better-timed follow-up study rather than claiming that increasing confidence has been proven to cause improved performance.

The case illustrates restrained interpretation. It does not report a real program effect or suggest that path analysis can replace experimental design.

Common misconceptions and mistakes

Path analysis is not proof of causation, and a significant indirect path does not automatically establish mediation in the causal sense. Good model fit cannot exclude every alternative explanation.

Standardized coefficients are not percentages. A path diagram can also conceal assumptions about residual correlations and measurement error. If variables are measured unreliably, treating them as error-free observed quantities can affect the estimates.

Business use

Businesses can use path analysis to make a proposed explanation explicit and examine whether observed data are consistent with parts of it. It is most useful when supported by a clear research question, credible measurement and suitable timing.

Specify the model before inspecting every possible path, and document alternatives and omitted influences. Report coefficients with uncertainty and explain whether they are standardized. Distinguish descriptive association, prediction and causal claims. If an operational decision depends on a mechanism being true, seek stronger evidence through design, longitudinal measurement or justified causal analysis rather than relying on diagram complexity.

Check your understanding

A fitted diagram has excellent overall fit. Does that prove the arrow from confidence to performance points in the correct causal direction?

Answer: No. Fit assesses compatibility with certain observed relationships, not the uniqueness of the causal explanation. Timing, design and alternative models still matter.

Related terms

Structural equation modeling
Mediation (statistics)
Confounding
mediation
direct effect
indirect effect
model identification
observed variable

Learn next

Learn next how model identification differs from model fit. Then write the equations behind a simple path diagram and list the assumptions that would be needed to interpret its indirect association as a causal effect.

Sources

Primary software documentation: lavaan mediation tutorial, https://www.lavaan.ugent.be/tutorial/mediation.html ; modelling tutorial, https://lavaan.ugent.be/tutorial/ . The lavaan mediation tutorial was consulted for linked equations, indirect and total paths, zero model degrees of freedom and uncertainty methods. Coefficient calculation and training-programme case are original illustrations.

Previous: Path Next: Path consistency learning