Statistics Glossary → Path Analysis
Path Analysisanalysis
An analysis that estimates a theoretical model of directed relationships among observed variables (who affects whom) as a system of simultaneous equations and tests how well the model fits the data.
Concept
Many research questions cannot be written as a single regression equation: a variable may be both a cause of one outcome and a result of another variable. Path analysis writes the arrow diagram (path diagram) that the researcher draws from theory as a system of equations and estimates all paths at once. Direct effects, indirect effects passing through intermediate variables and total effects are obtained together.
The model compares the observed covariance matrix of the variables with the covariance matrix implied by the drawn paths. If theory says some paths are zero (e.g. no direct arrow from X to Y), the model may not reproduce the data exactly; the gap is judged with the chi-square test and fit indices such as CFI, TLI, RMSEA and SRMR. An indirect effect is the product of the coefficients along the path. A 'saturated' model containing every possible path has zero degrees of freedom and fits perfectly by construction, so the fit test is informative only for models in which theory leaves some paths out.
Path analysis assumes that the variables are measured without error, that the model is correctly specified (no important variable or path omitted), and, for ML estimation, multivariate normality. Good fit does not prove the model is right; there can be models with different arrow directions that fit the same data equally well. The arrows assert a causal order, but that order comes from theory and design, not from the data themselves.
When to use it
- When a theory-based network of directed relationships made up of several equations (chained or parallel mediators, several outcomes) is to be tested with observed variables.
- If there is only one X, one or a few mediators and one Y, and the focus is the bootstrap interval of the indirect effect, mediation analysis is enough.
- If the variables are latent constructs measured by several items (e.g. the concept 'anxiety' rather than scale scores), structural equation modelling (SEM), which also models measurement error, is preferred.
An everyday example
A hospital management wants to understand how nurses' workload affects their job performance. Theory says workload raises stress, stress harms sleep quality and poor sleep lowers performance; workload may also have a direct effect on performance. Path analysis estimates these four variables and their arrows in one model, gives each direct and indirect effect, and uses the fit indices to show whether this chain fits the data from 250 nurses.
Parameters in MerQur
Template (Simple / Mediation / Serial / Parallel / Custom) and X, M, M1, M2, Y
Concept: Templates offer common path structures ready-made: simple (X → Y), mediation (X → M → Y), serial (X → M1 → M2 → Y) and parallel (from X to Y through two separate mediators). Each template sets up the equations defining which variable predicts which.
Everyday example: The hospital chain 'workload → stress → sleep → performance' fits the serial template. Had stress and sleep been thought of as two independent paths, the parallel template would be chosen.
What it does: Produces the model text: simple (Y ~ X), mediation (M ~ X; Y ~ X + M), serial (M1 ~ X; M2 ~ X + M1; Y ~ X + M1 + M2), parallel (M1 ~ X; M2 ~ X; Y ~ X + M1 + M2; no residual covariance between the mediators), custom (manual). A binary 0/1 X can also be selected.
Effect on the result: The whole model changes.
On the validation data: Mediation (hsbdemo): a = 0.7248 (SE 0.0580), b = 0.3654, c' = 0.4017, indirect effect 0.2649 (delta SE 0.0522). Parallel model: χ²(1) = 21.4258; CFI 0.9441, RMSEA 0.3196 [0.2110, 0.4436], SRMR 0.0638.
Model text (custom model)
Concept: In a custom model the paths are written as text: '~' states which variables predict a variable, and '~~' defines an undirected covariance (shared unexplained variation) between two variables.
Everyday example: The line 'performance ~ sleep + workload' says performance is affected by sleep and workload. 'stress ~~ burnout' accepts that the two vary together because of common causes not in the model.
What it does: "~" path, "~~" covariance (semopy syntax); written by hand in custom mode.
Effect on the result: The whole model changes.
On the validation data: Multiple independent variables (read ~ math + ses; science ~ read + math + ses): read ~ ses 1.726, science ~ ses 1.0337. Full mediation (science ~ read, c' = 0): χ²(1) = 28.8987.
Confidence Level (%90 / %95 / %99 / Other, custom value)
Concept: Sets the confidence level of the interval estimates for the path coefficients and the indirect and total effects. The RMSEA interval is conventionally reported at the 90% level.
Everyday example: If the 95% confidence interval for the effect of stress on sleep quality is [-0.55, -0.30], the effect is very probably clearly negative; choosing 99% widens the interval.
What it does: The level of the Wald interval and the bootstrap percentile interval of the path coefficients and the indirect/total effects. The RMSEA interval is always 90%.
Effect on the result: Only the intervals change.
On the validation data: read ~ math CI: 95% [0.6112, 0.8385], 90% [0.6294, 0.8202], 85.5% [0.6403, 0.8093].
Bootstrap (indirect effects) (0 = off)
Concept: Because indirect effects are products of coefficients, their distributions are skewed. The bootstrap draws many resamples from the data with replacement, recomputes the indirect effect in each and gives a confidence interval that does not rely on a formula.
Everyday example: The indirect effect of workload on performance through stress and sleep is the product of three coefficients. If the middle 95% of this product across 2,000 bootstrap samples excludes zero, the effect transmitted along the chain is significant.
What it does: Case resampling and a percentile interval for the indirect effects (seed fixed at 42).
Effect on the result: Only the indirect effect intervals change. Bootstrap is not done with FIML and a warning is written.
On the validation data: Mediation B = 500: [0.15339, 0.38552]; parallel B = 2000: [0.21255, 0.48089]; 90%: [0.23732, 0.45077].
Estimator (ML / MLM Satorra-Bentler / MLR Huber-White / GLS / WLS ADF)
Concept: ML is maximum likelihood estimation under multivariate normality. MLM and MLR scale the chi-square for departures from normality and give robust standard errors; GLS and WLS (ADF) estimate with different weightings, and WLS assumes no normality but needs a large sample.
Everyday example: If the nurses' stress scores are right-skewed, the ML chi-square may make fit look worse than it is. The scaled chi-square and robust standard errors of MLR give a more reliable result in the face of this skewness.
What it does: The β, SE, z, p and intervals of the path table are computed with the selected method; in a separate section the χ² of the method is given (scaled χ² and c in MLM and MLR). The standardized coefficients, R², effect decomposition and main fit indices are ML-based (stated with a note on screen).
Effect on the result: SE, z, p and χ² change.
On the validation data: Parallel model: MLM scaled χ² = 19.6876, c = 1.0883, read ~ math SE = 0.0549. MLR scaled χ² = 24.3466 (c = 0.8800). GLS χ² = 18.3521, WLS χ² = 18.7009.
Missing data (Listwise deletion / FIML)
Concept: Listwise deletion drops entirely any person missing on any variable. FIML (full information maximum likelihood) uses all of each person's observed values and gives unbiased results under the assumption that the missingness can be explained by the observed variables (MAR).
Everyday example: If 30 of the 250 nurses did not complete the sleep questionnaire, listwise deletion drops them completely and the sample falls to 220. FIML still uses these 30 nurses' workload, stress and performance data.
What it does: FIML uses all the observed data under the MAR assumption (only with ML and MLR).
Effect on the result: The sample size and all estimates change. If GLS is selected with FIML, a clear error message is given. GFI and AGFI are not computed in FIML.
On the validation data: boys: listwise n = 684: hgt ~ age 6.6169 (SE 0.0571), χ²(2) = 192.42. FIML n = 748: 6.5811 (SE 0.0549), χ²(2) = 211.708.
RMSEA denominator (N / N-1)
Concept: RMSEA measures the model's misfit per degree of freedom relative to sample size. Using N or N-1 in the formula creates small differences between software packages; with large samples the difference is negligible.
Everyday example: With 250 nurses, RMSEA coming out at 0.061 with N and 0.0612 with N-1 does not affect any decision. When comparing with another program, it is enough to make sure this setting matches.
What it does: The sample size used in the RMSEA calculation.
Effect on the result: Only the RMSEA point estimate changes; the 90% interval and PCLOSE are computed with N in both choices. It has no effect in a saturated model (df = 0).
On the validation data: Parallel model: 0.3196 with N; 0.3204 with N-1 (equal to the hand formula).
Mardia multivariate normality (Extra output)
Concept: Mardia's tests check whether the variables are jointly multivariate normal, in terms of skewness and kurtosis. A significant result suggests considering robust estimators such as MLM or MLR instead of ML.
Everyday example: If the multivariate skewness test is significant, a few nurses with both very high workload and very low sleep scores may form an extreme cluster. Estimating with MLR is then safer.
What it does: Multivariate skewness and kurtosis tests.
Effect on the result: Only adds output.
On the validation data: Parallel, 4 variables: b1p = 1.1426, χ²(20) = 38.087 (p = 0.0086); b2p = 23.3923, z = -0.6202. The coefficients use the covariance with an n denominator (ML).
Modification indices (Extra output)
Concept: A modification index shows roughly how much the chi-square would drop if a path or covariance not yet in the model were added. Only theoretically defensible suggestions should be considered; adding paths one after another because the data suggest them overfits the model to that sample.
Everyday example: The index may suggest adding a direct path from stress to performance. If it is theoretically reasonable that stress affects performance other than through sleep (e.g. via distraction), the path is added; otherwise the suggestion is ignored.
What it does: Lists the χ² decrease from adding a new parameter to the model (MI, EPC, sEPC). Candidates that would make an exogenous variable dependent are not listed.
Effect on the result: Only adds output.
On the validation data: Parallel model: read ~~ write MI = 20.318 (EPC 18.165, sEPC 0.3187).
Residual correlations (Extra output)
Concept: Residual correlations are the differences between the observed correlations and those implied by the model. A large residual shows the model fails to explain the relationship between those two variables.
Everyday example: If the observed correlation between workload and sleep is -0.40 and the model implies -0.20, the residual is -0.20: workload may affect sleep not only through stress but directly too, for example through night shifts.
What it does: Gives the matrix cor(S) - cor(Σ̂).
Effect on the result: Only adds output.
On the validation data: Parallel model: read-write residual correlation 0.188, write-science 0.053; the others |r| ≤ 0.030.
Comparison model
Concept: Two models are compared with the chi-square difference test (when one is a special case of the other, i.e. nested) and with the differences in AIC and BIC. A smaller AIC or BIC indicates a better balance between fit and simplicity.
Everyday example: Two models are compared, with and without the direct path from workload to performance. If the chi-square difference is not significant and BIC is smaller for the model without the path, the simpler explanation, that workload acts entirely through stress and sleep, is preferred.
What it does: Gives the chi-square difference, and the AIC and BIC difference, with an alternative model. The comparison is always done with ML and listwise deletion.
Effect on the result: Only adds output.
On the validation data: Parallel model versus saturated model (read ~~ write): Δχ² = 21.4258, Δdf = 1, ΔAIC = 19.43, ΔBIC = 16.13. Mediation versus full mediation: Δχ² = 28.8987.
Example on real data
Data: UCLA OARC hsbdemo
Variables: parallel template: X math, mediators read and write, Y science
- Question: Does the mathematics score (math) affect the science score (science) directly, or indirectly through reading (read) and writing (write)? (parallel path model with two mediators, n = 200)
- Paths: math strongly predicts read (β = 0.7248) and write (β = 0.6247); on science, the paths from math (0.3190), read (0.3015) and write (0.2065) are all significant (p ≤ .002).
- Indirect effect: The total indirect effect of math on science is 0.3476 (delta SE 0.0659, 95% CI [0.218, 0.477]); about half of the total effect of 0.6666 is transmitted through the mediators, and because the direct path also remains significant, the mediation is partial.
- Model fit: χ²(1) = 21.43, p < .001; CFI = 0.944, TLI = 0.665, RMSEA = 0.320 [0.211, 0.444], SRMR = 0.064. RMSEA and χ² show that the model does not fit the data; the overall fit is mediocre.
- Explained variance: R² is 0.4386 for read, 0.3812 for write and 0.4881 for science.
- Limitation: The parallel template assumes that the residuals of read and write are uncorrelated; the single degree of freedom comes from this constraint. The modification index suggests adding this covariance (this is the source of the poor fit).
- APA: Parallel mediation model, χ²(1) = 21.43, p < .001, CFI = .94, RMSEA = .32, SRMR = .06; total indirect effect = 0.35, 95% CI [0.22, 0.48].
Related concepts
Measurement model and structural modelModel fit indices (χ², CFI, TLI, RMSEA, SRMR)Modification indicesRobust ML estimators (MLM, MLR; Satorra-Bentler)Full information maximum likelihood (FIML)Indirect effect (mediation)Mardia's test of multivariate normalityPrediction error measures (RMSE, MAE, MAPE, MASE)
The numbers in the examples are MerQur's output on independent, publicly available validation data; see the MerQur Technical Note for details.