Written and maintained by CASRAI Editorial Board
Last updated
Path analysis is structural equation modeling with a structural model but no measurement model: every variable in a path diagram is a directly observed score, not a latent construct estimated from multiple indicators. That single fact explains most of what confuses researchers first learning it. The diagram looks like a full structural equation model (SEM)—boxes, arrows, standardized coefficients—because it is one, just a restricted case of it. What it estimates is a set of interrelated regression equations, solved so you can trace how a predictor’s total effect on an outcome splits into a direct piece and one or more indirect pieces running through other variables.
Path analysis was developed by geneticist Sewall Wright, who introduced it around 1918 and wrote about it more extensively through the 1920s, then formalized the path-tracing rules used to this day in a 1934 paper. It predates the latent-variable SEM most methods courses teach first by several decades, and it is still the right tool whenever your model has only observed variables—survey items used as scores rather than as indicators of a factor, administrative or biometric measurements, single-item ratings—and your real question is about the structure of direct and indirect effects among them, not about measurement.
What actually separates path analysis from full SEM
The distinction is not diagram style or software. It is whether the model includes a measurement component.
- Full SEM combines a measurement model (latent factors, each defined by multiple observed indicators, with factor loadings and indicator-level measurement error) with a structural model relating the latent factors to each other. See CASRAI’s guide to confirmatory factor analysis for how that measurement layer is built and evaluated on its own before it is ever combined with structural paths.
- Path analysis has only the structural layer. Every variable is a single observed score standing in for itself. There is no factor loading to estimate and no distinction between a construct and its indicator, because there is no construct—just the variable as measured.
The practical consequence: a path coefficient in path analysis is, definitionally, a standardized regression coefficient from one specific equation in a system of simultaneous regressions—the same quantity multiple regression reports for a single equation, just estimated jointly with the other equations in the system rather than one at a time. What path analysis adds over running separate regressions by hand is a single coherent model you can test for overall fit, and a formal accounting method for effects that pass through more than one variable.
Path diagram conventions
Because every variable is observed, a path diagram uses one shape throughout:
- Rectangles for every variable—there are no circles/ovals, which in a full SEM diagram are reserved for latent factors. If a diagram for your model has any circles, it is not a pure path model.
- Single-headed straight arrows for a hypothesized directional path, labeled with its standardized path coefficient (often written β or a lower-case letter like a, b, c′).
- Double-headed curved arrows for a correlation or covariance between two variables where neither is modeled as causing the other—used between exogenous (predictor-only) variables.
- A short unlabeled arrow into each endogenous (outcome) variable representing its disturbance (residual/error) term—the variance in that variable the model’s predictors don’t account for.
Worked example: tracing direct, indirect, and total effects
The numbers below are an illustrative worked example built to demonstrate the arithmetic, not a reported finding from any real study. Four observed variables:
X1 Supervisor support (survey score) X2 Peer collaboration frequency (survey score) M Research self-efficacy (survey score) Y Publication output, past 12 months (count) X1 <--- .30 ---> X2 | ___.45___ .33 | ___ / | .25 / | M | | | | .38 v v -----------------> Y (X1->M = .45, X2->M = .33, M->Y = .38, X1->Y = .25 direct. X2 has no direct path into Y.)
In words: X1 and X2 are exogenous and correlated (r = .30). Both have a path into M (X1→M = .45, X2→M = .33). M has a path into Y (M→Y = .38). X1 also has a direct path into Y (X1→Y = .25). X2 has no direct path into Y: the model specifies that peer collaboration affects publication output only by way of research self-efficacy, not directly.
| Path | Standardized coefficient | Interpretation |
|---|---|---|
| X1 → M | .45 | Direct effect of supervisor support on self-efficacy |
| X2 → M | .33 | Direct effect of peer collaboration on self-efficacy |
| M → Y | .38 | Direct effect of self-efficacy on publication output |
| X1 → Y | .25 | Direct effect of supervisor support on publication output |
| X1 ↔ X2 | .30 | Correlation between the two exogenous predictors (not a path) |
The tracing rule
Wright’s rule for combining coefficients along a chain: multiply the standardized coefficients along any legitimate chain (you may trace backward along an arrow and then forward, or straight through a sequence of forward arrows, but never forward and then back; you may pass through a given variable only once per chain; and at most one bidirectional correlation arrow may appear in a chain). Where more than one legitimate chain connects two variables, the total expected association is the sum of each chain’s product.
Decomposing the effect of X1 on Y
- Direct effect: X1→Y = .25
- Indirect effect (via M): X1→M × M→Y = .45 × .38 = .171
- Total effect: .25 + .171 = .421
About 41% of X1’s total effect on Y (.171 / .421) runs through research self-efficacy rather than directly—a decomposition a single regression of Y on X1 alone could never surface, because that regression’s coefficient on X1 would already be the .421 total, with no way to see the split.
Decomposing the effect of X2 on Y
X2 has no direct path to Y, so its entire effect on Y is indirect:
- Indirect effect (via M): X2→M × M→Y = .33 × .38 = .1254
- Total effect: .1254 (indirect only, since the direct path is fixed at zero by the model’s specification)
This is the same product-of-coefficients logic covered in more depth in CASRAI’s guide to mediation analysis, extended here to a model with two predictors and one mediator estimated simultaneously rather than a single X→M→Y chain estimated on its own. The same caution applies: the sampling distribution of a product of two coefficients is not normal even when each coefficient’s own distribution is, so a bias-corrected bootstrap confidence interval on the indirect effect is the current standard—not a Sobel test, and not the older causal-steps approach of testing each path’s significance in isolation and inferring mediation from the pattern. If your predictor and outcome aren’t clearly playing a mediator/moderator role in your hypothesis, CASRAI’s mediator vs. moderator comparison is the place to settle that before specifying the model.
Is the model identified, and can it actually be tested?
A recursive path model—every path runs in one direction, no feedback loops, and disturbance terms are not allowed to correlate with each other—can be estimated equation by equation with ordinary least squares, and its fit can be evaluated the same way a full SEM’s is (chi-square, CFI, RMSEA, SRMR), because a recursive path model is mathematically the special case of SEM where the measurement model is the identity: each latent variable has exactly one indicator with no measurement error. A non-recursive model, with reciprocal or looping paths, needs additional conditions (instrumental-type restrictions) to be estimable at all and is out of scope for an introductory model.
Whether a recursive model is even worth fitting for its own sake, rather than just read off as a set of regression results, depends on its degrees of freedom:
df = k(k+1)/2 − t, where k is the number of observed variables in the model and t is the number of freely estimated parameters (paths, exogenous variances and covariances, and endogenous disturbance variances).
Applied to the worked example above: k = 4 variables, so k(k+1)/2 = 10 known sample variances/covariances. Estimated parameters: 4 path coefficients (X1→M, X2→M, M→Y, X1→Y), 2 exogenous variances (X1, X2), 1 exogenous covariance (X1 with X2), and 2 disturbance variances (for M and for Y) = 9. df = 10 − 9 = 1.
That single degree of freedom is doing real work: it comes from the one path the model deliberately omits (X2→Y direct), which is exactly the testable claim that peer collaboration’s effect on output is fully mediated through self-efficacy, with nothing left over. A just-identified model (df = 0, a path or covariance drawn between every pair of variables) will always reproduce the observed covariance matrix perfectly by construction—a “perfect fit” that confirms nothing, because there was no way for it to fail. Testable fit requires at least one over-identifying restriction like this one; a saturated path diagram with every possible arrow drawn is not a hypothesis, it’s a re-description of the correlation matrix.
Reporting fit without over-claiming it
Because a recursive path model is fit and evaluated with the same machinery as full SEM, the same caution applies to reading its fit indices: CASRAI’s guide to SEM fit indices documents in detail why the popular “CFI ≥ .95, RMSEA ≤ .06, SRMR ≤ .08” combined cut-off is not actually what the Hu and Bentler (1999) paper it’s attributed to recommended, and why treating any single set of thresholds as a pass/fail gate overstates what the literature supports. That guidance applies unchanged to a path model with df > 0; it does not apply at all to a just-identified path model, since a model with no way to fit poorly has nothing for a fit index to test.
When path analysis is the right tool
| Your situation | Better tool |
|---|---|
| One predictor, one mediator, one outcome; you only need that single indirect effect and its confidence interval | Simple mediation via regression-based mediation (see the mediation analysis guide) — a full path model adds machinery you don’t need |
| Several observed predictors and outcomes, some effects hypothesized to be fully or partially indirect, and you want one testable model with an overall fit statistic | Path analysis |
| Your key variables are constructs measured by several items each (a scale, not a single score), and measurement quality itself is part of what you need to establish | Build and validate the measurement model first via confirmatory factor analysis (and, earlier still, exploratory factor analysis if the factor structure isn’t already established), then combine it with structural paths as full SEM |
Reporting checklist
- Report both standardized and unstandardized path coefficients with standard errors, not standardized coefficients alone—a reviewer checking your arithmetic needs the unstandardized values.
- Report the model’s degrees of freedom and, for any indirect or total effect you interpret, a bias-corrected bootstrap confidence interval rather than a significance test on the product of coefficients alone.
- State explicitly whether the model is recursive, and whether it is just-identified or over-identified—a just-identified model’s fit statistics are not evidence of anything and should not be reported as if they were.
- Report global fit indices (chi-square with its df, plus at least CFI and RMSEA or SRMR) only when df > 0, and interpret them as a pattern, not a single pass/fail threshold—see the SEM fit indices guide linked above for why.
- Name the software and version used (common choices: lavaan in R, Mplus, AMOS, or PROC CALIS in SAS) since minor differences in default estimator and standard-error method between packages can shift results at the margins.
Frequently asked questions
Is path analysis the same as structural equation modeling?
Path analysis is a form of SEM, not a separate technique—specifically, SEM with a structural model but no measurement model. Every variable in a path model is observed directly rather than estimated as a latent factor from multiple indicators. Once you add a measurement model (latent factors with multiple indicators each), you have moved from path analysis to full SEM.
Can path analysis include latent variables?
Not and still be called path analysis in the strict sense. The moment a variable is a latent factor estimated from indicators rather than a directly observed score, the model has a measurement component and is properly described as full SEM, even if parts of it look identical to a path diagram.
How do you calculate the total effect in path analysis?
Total effect equals the direct effect (if one is specified) plus the sum of every indirect effect, where each indirect effect is the product of the standardized coefficients along one legitimate chain connecting the two variables, per Wright’s tracing rule. If more than one distinct chain connects the same two variables, add each chain’s product into the total.
Is path analysis the same as mediation analysis?
They overlap but aren’t identical. Mediation analysis is usually a single X→M→Y question, often tested with regression or a dedicated macro. Path analysis is the more general framework: it can represent a single mediation relationship as a special case, but it’s built for models with several predictors, several outcomes, and several mediators estimated together as one system, with a testable overall fit statistic that a stand-alone mediation test doesn’t provide.
What software runs a path analysis?
Any SEM-capable package, since a recursive path model is a restricted SEM: lavaan (R, free and open source), Mplus, IBM SPSS Amos, and SAS’s PROC CALIS are the most commonly cited in published methods sections. A simple recursive model can also be estimated as a series of ordinary regressions by hand, though a dedicated package is what gives you the overall fit statistics.
