Path analysis shows how a set of measured variables affect each other in a chain, such as training leading to knowledge, knowledge to confidence and confidence to profit. This guide covers the ideas, the counting, the calculations, every panel of the output, a worked example and how to report it.
Path analysis maps how a set of variables affect each other, step by step. It is used when one thing leads to another, which leads to a third. A single regression cannot show this, because a regression has only one outcome. Path analysis fits several regressions together and shows how they connect.
The variables are things you measured directly, such as age, a test score or income. If your key variables are scores built from several questionnaire items, use a structural equation model instead.
A livelihoods programme trained 600 small business owners. Each person attended a number of training sessions. You also have a business knowledge score (0 to 100), a confidence score (0 to 100), monthly profit in US dollars, and age. You think training raises knowledge, knowledge raises confidence, and confidence and knowledge both raise profit. You want to know three things. Does each step in the chain hold? How much extra profit goes with one more session? And does the training reach profit only through knowledge and confidence?
Use it when you have a clear idea of a chain of influence among variables you measured. It suits programme theories with several steps, such as training, then skills, then practice, then results.
If you only have one step between a cause and an outcome, mediation analysis is simpler. If you have one outcome and many independent predictors, a linear regression is enough. If some variables are scores built from questionnaire items, use a structural equation model.
Path analysis can include yes or no outcomes and ordered outcomes as well as numeric ones. A sample of at least 200 is a sensible minimum, and the more arrows, the more cases you need.
| Point | What is needed | If it is not met |
|---|---|---|
| Measured variables | Every box in the model is a variable you measured, such as a score, an amount or a yes or no answer. | If a key variable is built from several questions, use a structural equation model so the noise in the questions is handled. |
| Sample size | At least 10 cases for every number estimated, and more for models with yes or no variables. The example estimates 12 numbers and uses 600 cases. | Estimates and intervals become unstable. |
| Direction | Arrows run one way, from earlier or causal variables to later ones. The data cannot confirm the direction. | A wrongly directed arrow can still fit. State your reasoning. |
| Variable types | Numeric, binary or ordered. Choose the right type for each. | The wrong type gives wrong numbers. A yes or no outcome treated as a number gives odd results. |
| Missing answers | Cases missing any variable in the model are left out. | Check missing values first. |
Figure 1 shows the example. Each box is a variable. Each arrow says that the variable at its start helps to explain the variable at its end, after taking the other arrows that lead to that variable into account. The numbers are standardised, so they can be compared.
Some variables are only causes, such as sessions and age. Some are only results, such as profit. Some are both, such as knowledge and confidence. Variables that are both are what make path analysis different from regression. They are called in-between variables.
The model for the example is three equations. Leaving an arrow out is a claim. Here there is no arrow from sessions straight to confidence or to profit, and none from age to confidence. The model says these effects go through the other variables.
That claim can be tested. The software compares the correlations the model predicts with the real ones. If an arrow you left out was needed, the gap shows in the fit. This is why path analysis has a fit test, even though each equation is simple.
The effect of one variable on another along a route is the product of the arrows on that route. If several routes lead to the same place, add them.
The effect of sessions on confidence has one route, through knowledge: 1.987 × 0.498 = 0.99. So one more session goes with about 0.99 extra points of confidence.
The effect of sessions on profit has two routes. The first goes through knowledge only: 1.987 × 0.500 = 0.99. The second goes through knowledge and then confidence: 1.987 × 0.498 × 1.045 = 1.04. Adding them gives 2.03. There is no direct arrow from sessions to profit, so the total is the same number.
The interval for a product of arrows is not a simple one, because a product of two uncertain numbers is more uncertain than either alone. When all variables are numbers, AnalyZ Solutions uses a standard approximation (the delta method) that works out the spread of the product from the spread of each arrow. When the final outcome is a yes or no variable, the module instead draws many new samples from your data, repeats the analysis on each, and takes the middle 95% of the results. You choose the number of resamples in that case.
Each variable that is led to by another gets its own equation. A variable that nothing leads to is a starting variable and has no equation.
All the equations are estimated together. That matters because a variable such as knowledge is an outcome in one equation and a predictor in another.
The example has 5 variables, so the data contain 5×6/2 = 15 variances and covariances. The model estimates 12 numbers: 6 arrows, 3 error variances for knowledge, confidence and profit, the variances of the 2 starting variables (sessions and age) and their 1 covariance. The difference, 15 - 12 = 3, is the degrees of freedom. The 3 comes from the arrows left out: sessions to confidence, sessions to profit and age to confidence. The fit tests whether leaving those arrows out is consistent with the data. If you draw an arrow between every pair of variables, the degrees of freedom are zero and there is no fit to test.
When all variables are numeric and no arrow loops back, each equation gives the same arrow estimates as a separate ordinary regression. What path analysis adds is the whole picture: it multiplies arrows along routes, gives the effects with intervals, and tests the arrows you left out. If your model has only one outcome, path analysis equals a regression.
| Your variables | Method | Intervals for indirect effects |
|---|---|---|
| All numeric | Maximum likelihood, all equations at once | A standard approximation (the delta method) |
| Final outcome yes or no | One probit regression for each equation | Resampling, with a number of resamples and a seed you can set |
| A mix of yes or no, ordered and numeric | Polychoric correlations with weighted least squares | From the survey-robust or model-based standard errors |
A probit regression is like a logistic regression, but it assumes a bell-shaped curve for the underlying tendency. Its numbers are on a different scale, so a probit arrow is not directly an odds ratio. Read its direction and size relative to other arrows in the same model.
In the example, sessions reach profit by two routes, 0.99 and 1.03, which add to 2.03 dollars of monthly profit per session. There is no direct arrow from sessions to profit, so the total is also 2.03.
A raw arrow is in the units of your variables, such as points of knowledge per session. A standardised arrow is in standard deviations, so you can compare a strong and a weak arrow even if the variables have different units. Use raw arrows to say how much, and standardised arrows to say which is stronger.
To test an arrow, compare the model with and without it. If leaving it out makes the chi-square clearly worse, keep it. AIC and BIC support the comparison: lower is better.
The fit shows whether the arrows you left out fit the data. Read it as in the other models in this module: CFI of 0.95 or higher, RMSEA of 0.06 or lower and SRMR of 0.08 or lower are good. If you include every possible arrow, the model fits perfectly by design, and there is nothing to test.
The table gives each arrow in its raw units, which say how many units the outcome changes for one unit of the cause, and in standardised form. It also gives an interval and a p-value. Raw values tell you what an effect means in practice. Standardised values tell you which links are strong.
For every pair of variables with more than one connection, the module separates the direct, indirect and total effect. It also lists each indirect route on its own, so you can see which one matters most.
For each outcome in the model, this is the share of its differences that the model accounts for.
You say how each variable is treated. Numeric variables are treated as numbers. Yes or no variables and ordered categories can be used as outcomes and predictors. Categorical predictors are turned into comparisons with a reference category, as in regression.
Open Structural Equation Modelling and choose Path Analysis. The module is marked Beta.
| Panel | What it shows | What to look for |
|---|---|---|
| Model fit | Chi-square, df, CFI, RMSEA, SRMR, AIC, BIC | Several measures in the good range. If df is zero, no fit is shown. |
| Path estimates | Raw and standardised arrows, with intervals and p-values | Direction, size and an interval away from zero. |
| Direct, indirect and total effects | For each pair of variables, the three effects with intervals | How much of an effect goes through the middle steps. |
| Routes | Each route with its own effect | Which route carries most of the effect. |
| Variance explained | Percent of each outcome explained | Low values mean other factors matter. |
| Interpretation | A plain-language summary | A first draft. Check it against the tables. |
The figures below are invented to show the steps. They come from a real run of the module on a simulated dataset of 600 business owners. Monthly profit averaged about 127 dollars.
| Chi-square | df | p-value | CFI | RMSEA | SRMR |
|---|---|---|---|---|---|
| 3.57 | 3 | 0.312 | 0.999 | 0.018 | 0.016 |
Step 1: check the fit. The chi-square is 3.57 on 3 degrees of freedom (p = 0.31), CFI is 0.999 and RMSEA is 0.018. The three arrows left out, sessions to confidence, sessions to profit and age to confidence, are not needed. The chain is a good description of the data.
| Path | Raw | Standardised | 95% CI | p-value |
|---|---|---|---|---|
| Training sessions → Business knowledge | 1.987 | 0.61 | 1.78 to 2.19 | < 0.001 |
| Age → Business knowledge | -0.106 | -0.10 | -0.18 to -0.04 | 0.003 |
| Business knowledge → Confidence | 0.498 | 0.49 | 0.43 to 0.57 | < 0.001 |
| Confidence → Monthly profit | 1.045 | 0.52 | 0.91 to 1.18 | < 0.001 |
| Business knowledge → Monthly profit | 0.500 | 0.24 | 0.36 to 0.64 | < 0.001 |
| Age → Monthly profit | -0.280 | -0.12 | -0.41 to -0.15 | < 0.001 |
Step 2: read the arrows. Each extra training session goes with 2.0 more points of business knowledge. Each extra knowledge point goes with 0.50 points of confidence. Each confidence point goes with about 1.05 dollars of monthly profit, and each knowledge point with about 0.50 dollars on top of that. Age has small negative links with knowledge and profit. All arrows are clear, since none has an interval that includes zero.
| Effect | Direct | Indirect | Total | 95% CI of indirect |
|---|---|---|---|---|
| Sessions → Confidence | 0.00 | 0.99 | 0.99 | 0.82 to 1.16 |
| Sessions → Profit | 0.00 | 2.03 | 2.03 | 1.68 to 2.38 |
| Knowledge → Profit | 0.50 | 0.52 | 1.02 | 0.42 to 0.62 |
Step 3: read the effects. One more session goes with about 2.0 dollars more profit per month (interval 1.7 to 2.4). All of it runs through knowledge, and about half of it also through confidence. Since the model has no direct arrow from sessions to profit, the whole effect is indirect. 1.04 dollars of it runs through knowledge and then confidence, and 0.99 through knowledge alone.
| Outcome | Explained by the model |
|---|---|
| Business knowledge | 38% |
| Confidence | 24% |
| Monthly profit | 48% |
Step 4: read the variance explained. The model accounts for about 38% of the differences in knowledge, 24% in confidence and 48% in profit. A large part of each remains unexplained, which is normal.
Business owners who attended more sessions knew more about running a business. Those who knew more felt more confident, and those who knew more and felt more confident earned more. The training appears to raise profit by building knowledge and confidence, with no sign of a separate route. The model describes patterns in this sample. Owners who chose to attend more sessions may differ in ways that were not measured, so this is not proof that training causes the gain.
Treating the arrows as proven causes. They are your theory. The data can show whether the theory fits, not that it is true.
Ignoring the order in time. A cause has to come before its result. If variables were all measured at the same moment, say that in your report.
Leaving out an important variable. If something affects both the cause and the result and you did not measure it, the arrow between them can be misleading.
Adding every possible arrow. The model then fits by design and tests nothing.
Reading standardised values as amounts. Use raw values to say what an effect means in dollars, points or days.
Too many arrows for the sample. Simplify when the model will not run or gives odd values.
Open Structural Equation Modelling in AnalyZ Solutions. Free to sign up, free to analyze.
Get started