What Is SEM? How to Read SEM Results in SmartPLS
What is SEM
SEM, short for Structural Equation Modeling, is a group of methods used to test relationships between research constructs at the same time. In an SEM model, you can assess both the measurement model, where items measure latent variables, and the structural model, where paths represent effects between those latent variables.
Suppose your research hypothesizes that Service Quality affects Satisfaction, and that Satisfaction then affects Repurchase Intention. If each construct is measured with several items, ordinary regression does not fully address measurement error. SEM allows you to model latent variables and test the full hypothesized structure in one procedure.
An SEM model usually has two parts:
- The measurement model, also called the measurement model, assesses outer loading, Cronbach's Alpha, Composite Reliability, AVE, and HTMT.
- The structural model, also called the structural model, assesses path coefficient, p-value, t-value, R², f², Q², and VIF.
You can review the basic concepts in the SEM model. When you open a SmartPLS model, first identify whether the statistic belongs to the measurement model or the structural model. Many interpretation errors happen when a measurement model statistic is used to draw a conclusion about a hypothesis.
In general form, a structural equation can be written as follows:
Y = β1X1 + β2X2 + ζ
Here, Y is the dependent variable, X1 and X2 are predictors, β is the path coefficient, and ζ is the error term. For a model with latent variables, SmartPLS also estimates the relationships between each latent variable and its items.
Why SEM matters in quantitative research
SEM helps you answer a chain of research questions instead of testing only separate pairs of variables. For example, you can determine whether Trust affects Usage Intention, how strong that effect is, and whether Perceived Usefulness mediates the relationship between the two variables.
The important point is that SEM does not create a theory for your thesis. You still need to define the conceptual model, hypotheses, scale, and expected direction of effects before you run the analysis. The output tells you whether the data support the hypotheses you specified.
In quantitative research, SEM is commonly used for four decisions:
- Assess whether the scale measures the intended construct.
- Assess whether the latent variables are distinct from one another.
- Assess the direction and strength of effects between variables.
- Assess how much variance in the dependent variable the model explains.
If you use SmartPLS, the usual approach is PLS-SEM. PLS-SEM is suitable when the objective focuses on prediction, the model contains many latent variables, or the data do not meet distributional assumptions particularly well. You can review the different procedures in PLS-SEM.
CB-SEM is another approach. It is commonly associated with evaluating overall model fit using indices such as CFI, TLI, RMSEA, and SRMR. These indices belong to the CB-SEM context and should not be inserted mechanically into a SmartPLS results discussion. If you are deciding between the two approaches, What is CB-SEM separates the two procedures clearly.
How much SEM is acceptable
The question "how much SEM is good" needs to be separated into groups of statistics. There is no single number that determines whether the entire SEM model is acceptable. A model can have a high R² while still having an HTMT problem. A path coefficient can be statistically significant while the scale remains unreliable, which is still insufficient for defending the result.
The table below uses thresholds commonly reported in PLS-SEM. Read each threshold together with the research objective, the type of scale, and the context of the data.
| Assessment group | Statistic and reference threshold | Interpretation | Source |
|---|---|---|---|
| Reliability | Cronbach's Alpha of 0.7 or above | The scale has acceptable internal consistency | (Nunnally, 1978) |
| Exploratory reliability | Alpha from 0.6 may be acceptable | May be considered for a new scale or exploratory research | (Hair et al., 2010) |
| Composite reliability | CR of 0.7 or above | The items have suitable composite reliability | (Fornell and Larcker, 1981) |
| Convergent validity | AVE of 0.5 or above | The latent variable explains sufficient variance in its items | (Fornell and Larcker, 1981) |
| Outer loading | 0.7 or above | The item contributes well to the latent variable | (Chin, 1998) |
| Discriminant validity | HTMT below 0.85, or below 0.90 for closely related constructs | The two latent variables can be distinguished | (Henseler et al., 2015) |
| Multicollinearity | VIF below 5 | No serious multicollinearity is indicated under the reference threshold | (Hair et al., 2019) |
| Explanatory power | R² of 0.75, 0.50, and 0.25 are substantial, moderate, and weak, respectively | Assesses the model's ability to explain the dependent variable in PLS-SEM | (Hair et al., 2011) |
| Effect size | f² of 0.02, 0.15, and 0.35 are small, medium, and large, respectively | Assesses the contribution of one predictor | (Cohen, 1988) |
| Predictive relevance | Q² greater than 0 | The model has predictive relevance | (Stone, 1974) |
A path coefficient has no single threshold that is good for every research topic. Check whether the coefficient has the hypothesized direction, whether the p-value or bootstrap confidence interval indicates an effect different from 0, and then assess practical significance based on the size of the coefficient.
SmartPLS commonly uses bootstrapping to test the significance of a path coefficient. Under the current PLS-SEM procedure, 5,000 bootstrap subsamples are commonly reported (Hair et al., 2022). This number is a setting for the estimation procedure, not a criterion for deciding whether a hypothesis is accepted.
How to read SEM results in SmartPLS
After running PLS-SEM, read the results in order, starting with the measurement model and then moving to the structural model. Do not begin by looking at the p-value for H1 while ignoring outer loading, CR, AVE, and HTMT.
The table below is illustrative output, arranged in the format of SmartPLS output. It is not the result of a real study.
| SmartPLS output area | Statistic | Illustrative value | Interpretation |
|---|---|---|---|
| Outer Loadings | CLDV1 | 0.781 | The item meets the reference threshold of 0.7 |
| Construct Reliability and Validity | Cronbach's Alpha, CLDV | 0.842 | Internal reliability is acceptable |
| Construct Reliability and Validity | Composite Reliability, CLDV | 0.889 | Composite reliability is acceptable |
| Construct Reliability and Validity | Average Variance Extracted, CLDV | 0.667 | Convergent validity is acceptable |
| Discriminant Validity | HTMT, CLDV and HL | 0.812 | Below 0.85, meeting the stricter threshold |
| Collinearity Statistics, VIF | CLDV → HL | 2.146 | Below 5, with no current indication of a multicollinearity problem |
| Path Coefficients | CLDV → HL | 0.436 | Positive effect, with the magnitude interpreted in context |
| T Statistics | CLDV → HL | 5.274 | Read together with the p-value or bootstrap confidence interval |
| P Values | CLDV → HL | 0.000 | Significant at the selected significance level, but do not report p as exactly 0 |
| R Square | HL | 0.528 | The model explains 52.8% of the variance in HL |
| f Square | CLDV → HL | 0.241 | The illustrative effect size is medium |
| Q Square | HL | 0.316 | Greater than 0, indicating predictive relevance |
In the Outer Loadings table, check whether each item loads on the correct latent variable. If a loading is low, assess it together with the scale content, CR, and AVE. Do not immediately delete a low-loading item simply to make the results table look better.
In Construct Reliability and Validity, read Alpha, rho_A, Composite Reliability, and AVE for each latent variable. If Alpha meets the threshold but AVE is low, the scale may still lack convergent validity. If CR is extremely high, check whether the items are repeating almost exactly the same content.
In Discriminant Validity, HTMT requires close attention. If HTMT exceeds the threshold, review the conceptual distinction between the two variables instead of looking only at the number. You can also read what a latent variable is to distinguish latent variables, items, and the relationships between them.
In Path Coefficients, a positive or negative sign indicates the direction of the effect. The T Statistics and P Values columns, together with the bootstrap confidence interval, help you assess statistical significance. R Square belongs to the endogenous variable, while f Square indicates how much a predictor contributes to the dependent variable.
What to do when SEM does not meet the thresholds
When a statistic does not meet the threshold, work through the possible causes in an order that you can explain. First, check the original data, variable names, scale coding, and reverse-coded items. A reverse-coded item that has not been recoded can produce a low loading and reduce the reliability of the entire scale.
Check data and model errors
Review the .csv file, delimiter, blank cells, data types, and whether each item has been assigned to the correct latent variable. Then check whether the model contains an arrow pointing in the wrong direction, a variable connected incorrectly, or a hypothesized relationship that was left out.
Consider removing an item
If an outer loading is low, you may consider removing the item after checking its theoretical meaning and its effect on CR and AVE. Remove only one item at a time when there is a clear reason, rerun the model, and record the change. Deleting items in bulk based only on the number can move the scale away from its original source.
Reassess discriminant validity
A high HTMT may indicate that two constructs overlap. Compare the content of the items and check whether any item has almost the same meaning as another, or whether respondents understood the two constructs in the same way. Combining two latent variables is reasonable only when there is a theoretical basis and your supervisor approves it.
Reassess the structural model
If VIF is high, check whether the predictors are too similar. If R² is low, do not automatically add variables simply to increase R². A compact model that fits the theory and has reasonable explanatory power is usually easier to defend than a model with too many paths.
Some problems cannot be repaired by software because they come from the questionnaire design or the collected data. If many respondents selected the same response option, the data have little variation, or the missing-data rate is too high, discuss data cleaning, additional data collection, or limits on the conclusions. Do not create additional rows or edit responses to force the SEM model to meet the thresholds.
Distinguishing SEM, PLS-SEM, and CB-SEM
SEM is the broad name for a group of models. PLS-SEM and CB-SEM are two common approaches within that group. SmartPLS is mainly used to run PLS-SEM, while AMOS is commonly used for CB-SEM. SPSS can support regression, scale assessment, and some mediation procedures, but it does not fully replace dedicated SEM software.
PLS-SEM focuses strongly on explained variance and prediction. You will commonly see the Outer Loadings, Path Coefficients, R Square, f Square, Q Square, and HTMT tables. CB-SEM focuses on the fit between the observed covariance matrix and the theoretical model. Therefore, CFI, TLI, and RMSEA need to be interpreted in the CB-SEM context.
SEM is also easy to confuse with multivariate regression. Regression often uses observed variables or precomputed composite scores, while SEM allows you to model latent variables with multiple indicators and test several equations at the same time. If your topic has only one dependent variable, a small number of predictors, and no mediation structure, regression may be the more compact choice.
You should also distinguish SEM from mediation and moderation. Mediation is a relationship structure inside SEM in which a mediator transmits part or all of an effect. Moderation tests whether one variable changes the strength or direction of the relationship between two other variables. Neither term is a replacement name for SEM.
Common mistakes when reading SEM
The first mistake is concluding that "SEM is acceptable" because a few p-values are small. A p-value answers a question about the statistical significance of a relationship. You still need to assess scale quality, discriminant validity, VIF, R², and predictive relevance.
The second mistake is using CB-SEM thresholds for SmartPLS. CFI, TLI, and RMSEA are not the default criteria for PLS-SEM. If your thesis uses SmartPLS, report the appropriate PLS-SEM statistics and state clearly which procedure you ran.
The third mistake is writing p = 0.000. Software often displays three decimal places, but that does not mean the probability is exactly 0. The appropriate format is p < 0.001 when the output displays 0.000.
The fourth mistake is deleting items until every statistic looks good. Each item must have substantive meaning and a scale source. Save the different model versions, record why each item was removed, and report the procedure in Chapter 4.
The final mistake is mixing illustrative values with real study results. The example table helps you identify column locations and understand the interpretation. When writing your thesis, replace it with output from your own data file and check the number of observations, variable codes, and hypothesized paths.
Frequently asked questions
What is SEM in quantitative research?
SEM is structural equation modeling, used to test the measurement model and structural model at the same time. It is suitable when a study has several latent variables, each measured by multiple items, and hypothesized relationships that need to be tested together.
How much SEM is good?
There is no single number for the whole SEM model. Review each group of statistics, including outer loading, CR, AVE, HTMT, VIF, R², f², Q², and the significance of the path coefficient. Thresholds are sourced reference points, while the final conclusion must be based on the specific model and data.
Should I run SEM with SmartPLS or SPSS?
SmartPLS is suitable for running PLS-SEM with latent variables, bootstrapping, and statistics such as HTMT, R², f², and Q². SPSS is more suitable for data cleaning, descriptive statistics, Cronbach's Alpha, EFA, correlation, and regression. You can use both programs at different stages if the method specified for your study permits it.
What should I do if SEM does not meet the thresholds?
Check the data, variable coding, reverse-coded items, model structure, and item assignments first. Only then consider removing a low-loading item or adjusting the model, provided that you have a theoretical reason and record each run. If the problem comes from limited variation in the data or a weak questionnaire, software cannot replace a review of the data.
What R² is good in SEM?
In PLS-SEM, R² values of 0.75, 0.50, and 0.25 are commonly interpreted as substantial, moderate, and weak, respectively (Hair et al., 2011). These are reference levels rather than mandatory standards for every field. Read R² together with the path coefficient magnitude, f², Q², and the model's prediction objective.
Do I need to report every SEM statistic in my thesis?
Report the statistics that match the procedure you registered, including the measurement model, structural model, and hypothesis tests. You do not need to place every raw output table in the main text if doing so makes Chapter 4 difficult to read, but keep the files so you can explain them when asked.
Open your SmartPLS file now, separate the results into the measurement model and structural model, and compare each statistic with the sources in the table above. If you need to run and interpret your own .sav or .csv data, see M4 data analysis.