A structural equation model measures topics such as quality or trust from several survey questions and tests how the topics affect each other, all in one model. This guide covers the ideas, the counting, the calculations, every panel of the output, a worked example and how to report it.
A structural equation model tests how topics that you measure with several questions affect each other. It joins two ideas. The first is measuring a hidden topic, such as trust, from a set of questions. The second is asking which topic influences which.
It is often shortened to SEM. The name sounds heavy, but the question it answers is plain: if one topic goes up, what happens to the others, and through which route?
You surveyed 500 people in a host community. Eleven questions cover three topics: service quality (four questions), trust in providers (four questions) and use of services (three questions, such as how likely people are to go back). You think good service builds trust, and trust makes people use the services. You want to know three things. Does quality raise use? Does it do so only by building trust, or also directly? And how much of the differences in trust and use do these topics explain?
You can, and it is a fine first look. But an average of questions still carries the noise of each question. That noise weakens the links between topics, so the effects look smaller than they are. A structural equation model separates the topic from the noise before it estimates the links. It also lets one topic be the result of one relationship and the cause in the next, all in one model.
Use it when your key variables are scores built from several questionnaire items and you have a clear idea of how they connect. Typical cases are attitudes, trust, satisfaction, well-being and similar topics.
If all your variables are single measured numbers, such as age, income or a test score, use path analysis. If you only want to know whether one variable carries the effect of another, mediation analysis is simpler. If you only want to check that questions group as planned, use confirmatory factor analysis.
You need a model idea before you start, and a sample that is large enough. The module asks for at least 200 cases, and more for bigger models.
| Point | What is needed | If it is not met |
|---|---|---|
| Sample size | At least 200 cases, and more when the model has many topics, few questions per topic or weak loadings. As a floor, 10 cases for every number the model estimates. | Estimates and fit become unstable. Standard errors are too wide to be useful. |
| Topics and questions | Each topic has at least two questions, and three or more is safer. Each question belongs to one topic. | The module stops with a message. |
| A model idea | The arrows come from theory or from the design of the programme, not from what the data happen to show. | The result describes a model built to fit, and will probably not hold on new data. |
| Direction of arrows | Each arrow says one thing leads to another. The data cannot confirm the direction. Only your design and logic can. | A model with the arrows reversed often fits equally well. |
| No loops | The module fits models where effects run one way. Two topics cannot each affect the other. | The module stops, or the model cannot be estimated. |
| Reverse-worded questions | Recode first, so a high score always means more of the topic. | Negative loadings and poor reliability. |
| Missing answers | A case missing any variable in the model is left out. The cases used are shown. | Check missing values first if many cases drop out. |
A structural equation model has a measurement part and a structural part.
Figure 1 shows the structural part of the example. Solid arrows are links you expect. The dashed arrow is a direct link from quality to use that skips trust. You include it to find out whether trust really carries the whole effect.
The numbers on the arrows are standardised, so they run roughly from -1 to 1 and can be compared. A value of 0.60 means that people who are one standard deviation higher on quality are about 0.6 of a standard deviation higher on trust.
The software estimates the measurement part and the structural part at once. As in confirmatory factor analysis, it picks the loadings, error sizes and path values that make the correlations predicted by the model as close as possible to the correlations in your data. Estimating together is what lets it separate noise from the topics.
An effect can travel in more than one way. Quality can reach use directly, and it can reach use through trust. The indirect effect is the product of the arrows along the route. The total effect adds the direct and indirect parts.
With the unstandardised values from the example: 0.686 × 0.436 = 0.299 for the indirect effect. The direct effect is 0.063, so the total is 0.063 + 0.299 = 0.362. In standardised terms the indirect effect is 0.599 × 0.558 = 0.33.
Trust is predicted only by quality, so the variance explained is the square of the standardised path: 0.599² = 0.36. About 36% of the differences in trust are accounted for by quality. For use, quality and trust together account for about 36%.
To test whether a path is needed, fit the model with and without it and compare the chi-square values. Without the direct path, the model fits with a chi-square of 24.3 on 41 degrees of freedom. With it, the chi-square is 23.3 on 40. The difference is 1.0 for one extra path, which is small. Adding the direct path barely improves the model.
A structural equation model is two models joined together. The measurement part says each question is a noisy reading of a topic. The structural part says how the topics affect each other.
The topics are never observed. The model finds the topic values that best explain the answers, and it estimates the loadings and the paths at the same time. This is why the paths are not weakened by the noise in the questions. If you average the questions and run a regression, the noise in the averages pulls the paths toward zero.
The example has 11 questions: 4 on quality, 4 on trust and 3 on use. That gives 11×12/2 = 66 variances and covariances for the model to explain. The model estimates 26 numbers: 8 loadings (the first question of each topic is fixed to set the scale), 11 error variances, 3 paths, 1 variance for the quality topic, 2 residual variances for trust and use, and 1 correlation between the errors of two quality questions (the change suggested in the confirmatory analysis). The difference, 66 - 26 = 40, is the degrees of freedom. It counts the independent ways the model can disagree with the data. The paths you leave out, and the questions you do not let load on other topics, are what make the model testable. A model in which every topic is linked to every other, and every question to every topic, would fit perfectly and tell you nothing.
If a model estimates exactly as many numbers as there are variances and covariances, the degrees of freedom are zero and the fit is perfect by design. If it estimates more, it cannot be estimated, and the module tells you to simplify it.
Estimation works as in confirmatory factor analysis. For numeric questions the module uses maximum likelihood. It searches for the loadings and paths that make the variances and covariances implied by the model as close as possible to the real ones, and the chi-square is the leftover gap multiplied by the number of cases. For ordered or yes or no questions it uses polychoric correlations and weighted least squares with an adjusted chi-square. With a survey weight, estimates use the weights and standard errors follow the survey design.
When a topic affects another through a middle topic, the indirect effect is the product of the arrows along the route. The total effect is the direct effect plus all indirect effects.
In the example the standardised paths are 0.60 and 0.56, so a rough check gives 0.60 × 0.56 = 0.34. The table shows the exact indirect effect, which is worked out from the unrounded estimates. The direct path from Quality to Use is 0.07 and its interval includes zero, so the total effect of 0.362 is almost all indirect. Intervals for these products are worked out with the delta method for numeric variables.
The fit measures are those used in confirmatory factor analysis: chi-square, CFI, TLI, RMSEA with its range, and SRMR. Here the model has chi-square 23.3 on 40 degrees of freedom, CFI 1.000, RMSEA 0.000 and SRMR 0.020. Good fit has CFI and TLI of 0.95 or more, RMSEA of 0.06 or less, and SRMR of 0.08 or less. Read all of them together.
For each outcome topic, the module gives the share of its variation that the topics leading to it account for, as a percentage. It tells you how much your model explains, not whether it is right.
To test whether an arrow is needed, run the model with and without it. The model with fewer arrows has more degrees of freedom. If its chi-square is not much worse, the simpler model is enough. The AIC and BIC columns help compare models that are not nested. Lower values are better, and BIC penalises extra arrows more strongly.
Read the fit as in a confirmatory factor analysis. CFI and TLI of 0.95 or higher, RMSEA of 0.06 or lower and SRMR of 0.08 or lower are good. A model that fits badly cannot be trusted, even when its paths look strong. Check fit first.
The measurement part is reported as in confirmatory factor analysis: standardised loadings, composite reliability, average variance extracted, and the heterotrait-monotrait ratio for separation. A weak measurement part makes the structural results hard to trust.
The paths table gives each link with its raw value, standardised value, interval and p-value. An interval that includes zero means the link could be zero. Standardised values let you compare links within the same model.
For each pair of topics with more than one route, the module separates the direct, indirect and total effect, with intervals. It also lists each indirect route on its own, so you can see which one carries the effect.
This shows how much of each outcome the model accounts for. There is no universal good value. It depends on the field.
As in confirmatory factor analysis, the module suggests changes that would improve fit. Add one only if you can explain it in plain words.
For items that are yes or no, or only a few ordered categories, choose the mix of binary, ordinal and continuous. Predictors that are not factors, such as sex or age group, can be added as ordinary variables, and you say whether each is numeric, binary or categorical.
Open Structural Equation Modelling and choose Structural Equation Model. The module is marked Beta.
| Panel | What it shows | What to look for |
|---|---|---|
| Model fit | Chi-square, df, CFI, TLI, RMSEA with its range, SRMR, AIC, BIC. | Several measures in the good range, not one. |
| Loadings, reliability and separation | As in confirmatory factor analysis: loadings, composite reliability, AVE, factor correlations, discriminant validity. | Fix the measurement part first. Weak loadings weaken everything built on them. |
| Structural paths | Each arrow between topics or variables, raw and standardised, with intervals and p-values. | The direction, size and interval of each arrow. An interval that includes zero means no clear effect. |
| Effects route by route | Direct, indirect and total effects, route by route, with intervals. | Whether the effect goes mostly through the middle topic. |
| Variance explained | The share of each outcome explained. | Low values mean other things matter more than the model includes. |
| Suggested changes | Modification indices for paths and covariances left out. | Add a change only if you can explain it in words. |
| Interpretation | A plain-language summary. | A first draft. Check it against the tables. |
| Message | What it means | What to do |
|---|---|---|
| The model is not identified | The data do not contain enough information for all the numbers. | Add questions to topics with only two, or remove arrows. |
| More estimates than the data can support | The model has more numbers than the data have variances and covariances. | Simplify the model. |
| An error variance is at or near zero | A question is almost the same as its topic. | Check the question. Add a third question to the topic. |
| A standardised loading is above 1 | An improper solution, often from a small sample or too few questions. | Simplify the model. Check for near-duplicate questions. |
| The model did not fully converge | The search did not finish cleanly. | Treat results with caution. Simplify or add cases. |
The figures below are invented to show the steps. They come from a real run of the module on a simulated survey of 500 people, with the questions described earlier.
| Chi-square | df | p-value | CFI | TLI | RMSEA (90% CI) | SRMR |
|---|---|---|---|---|---|---|
| 23.3 | 40 | 0.984 | 1.000 | 1.011 | 0.000 (0.000 to 0.000) | 0.020 |
Step 1: check the fit. CFI is 1.000, RMSEA is 0.000 and SRMR is 0.020, so the model fits the data well. A fit this clean is unusual in real surveys. With real data, expect some compromise.
| Factor | Composite reliability | AVE | Weakest loading |
|---|---|---|---|
| Service quality | 0.79 | 0.48 | 0.66 |
| Trust in providers | 0.82 | 0.54 | 0.55 |
| Use of services | 0.75 | 0.50 | 0.69 |
Step 2: check the measurement part. Composite reliability is 0.75 or higher for every factor, and the weakest single loading is 0.55. The topics are measured well enough to interpret the paths.
| Path | Raw | Standardised | 95% CI | p-value |
|---|---|---|---|---|
| Service quality → Trust in providers | 0.686 | 0.60 | 0.55 to 0.82 | < 0.001 |
| Trust in providers → Use of services | 0.436 | 0.56 | 0.32 to 0.55 | < 0.001 |
| Service quality → Use of services (direct) | 0.063 | 0.07 | -0.06 to 0.19 | 0.321 |
Step 3: read the paths. Quality has a clear effect on trust (standardised 0.60). Trust has a clear effect on use (0.56). The direct path from quality to use is small (0.07) and its interval, -0.06 to 0.19, includes zero.
| Effect | Raw | 95% CI | Standardised |
|---|---|---|---|
| Direct | 0.063 | -0.06 to 0.19 | 0.07 |
| Indirect, through trust | 0.299 | 0.20 to 0.39 | 0.33 |
| Total | 0.362 | 0.25 to 0.47 | 0.41 |
Step 4: read the effects. The total effect of quality on use is 0.36. Almost all of it, 0.30, goes through trust. The direct part is small and cannot be told from zero. Roughly 83% of the total effect travels through trust.
Step 5: look at variance explained. Quality accounts for about 36% of the differences in trust. Quality and trust together account for about 36% of the differences in use.
In this survey, people who rated services as better trusted providers more, and people who trusted providers more said they would use the services more. Trust appears to carry nearly all of the link between service quality and use. Once trust is taken into account, quality adds almost nothing on its own. The model describes patterns in this sample. It does not prove that improving quality would raise trust.
Reading arrows as proof of cause. The model shows patterns that agree or disagree with your idea. Data collected at one point in time cannot show which way influence runs.
Skipping the measurement check. If a topic is poorly measured, every path that involves it is unreliable.
Building a model with no theory. Trying many arrow patterns until one fits produces a model that rarely holds up.
Too complex a model for the sample. More factors and paths need more cases. If the model will not converge or gives odd values, simplify it.
Ignoring other models that fit equally well. A model that fits does not rule out others that fit as well. Say so in your report.
Judging a path by its p-value alone. Look at its size and interval as well.
Open Structural Equation Modelling in AnalyZ Solutions. Free to sign up, free to analyze.
Get started