Every thesis season, one question flows to me almost nonstop: “Should I use LISREL or SmartPLS?” And almost always, behind that question is the same reason: “My friend says SmartPLS is easier.”
I understand. As someone who has written teaching modules on both approaches, I will not deny that SmartPLS is friendly on screen. But choosing an analysis method because of “whichever looks friendlier” is like choosing a vehicle by its color. It may be pleasant, but it does not mean you will reach your destination.
This article tries to offer a healthier way of thinking about choosing between two families of Structural Equation Modeling (SEM): covariance-based SEM (CB-SEM) and variance-based SEM (PLS-SEM).
First, what is SEM?
SEM is a single analytical framework that can answer several questions at once. It describes the relationships between latent variables (concepts that cannot be measured directly, such as satisfaction, motivation, or trust) and their indicators, as well as the relationships among the latent variables themselves. The main advantage of SEM over ordinary regression is that it accounts for measurement error and analyzes many relationships in one model.
One thing I always remind people: SEM is not just a technique, it is a way of thinking. You start from theory, draw the model, and only then test whether the data supports the theory. If you reverse that order (run the software first, look for theory afterwards), the results are usually a mess.
Two families, two philosophies
CB-SEM (for example LISREL, AMOS, Mplus, or the lavaan package in R) tries to reproduce the covariance matrix of the data. The main question: does my theoretical model fit the pattern of relationships in the data? That is why it produces many goodness-of-fit measures such as RMSEA, CFI, TLI, and SRMR. CB-SEM suits testing and confirming theory.
PLS-SEM (for example SmartPLS) works with a variance-based approach. It tries to maximize the explained variance of the dependent variable. The main question: how well does this model predict and explain the variables I have linked? PLS-SEM suits theory development and prediction.
One philosophical difference, two ways of judging quality. That is the root of every technical difference below.
When does CB-SEM make more sense?
- Your main goal is to test a well-established theory, and you want to know whether the whole model fits the data.
- Your constructs are mostly reflective (indicators are seen as reflections of the construct).
- Your sample is large enough and the data is close to the required distributional assumptions (or you use an estimator that is more robust to violations).
- You need advanced features such as testing measurement invariance across groups.
When does PLS-SEM make more sense?
- Your main goal is prediction or exploration of a model whose theory is not yet mature.
- Your model is complex, with many constructs and indicators.
- You have formative constructs (indicators “form” the construct rather than being reflected by it).
- Your data does not meet normality assumptions, or the sample is relatively small, although this point is often misunderstood (see below).
Three myths to clear up
Myth 1: “PLS-SEM suits small samples, so any sample size is safe.” PLS-SEM is indeed more tolerant than CB-SEM of small samples and non-normal data, but that does not make it magic. Small samples still produce unstable estimates. The often-quoted “ten times” rule is criticized for being too crude. Sample requirements still depend on model complexity and the effect size you want to detect.
Myth 2: “PLS-SEM has no fit measures, so it is easier to pass.” The measures are different, not absent. In PLS-SEM we assess the measurement model through outer loadings (usually ≥ 0.70 as a benchmark), composite reliability, average variance extracted (AVE ≥ 0.50), and discriminant validity (for example HTMT). The structural model is assessed through R², effect sizes, predictive relevance (Q²), and the significance of path coefficients via bootstrapping. The evaluation is no looser. It just speaks a different “language”.
Myth 3: “One approach is certainly superior.” There is no such thing. Both are tools, and both can be used wrongly. Many studies reach substantively similar results when the model and data meet the requirements of both approaches.
Guiding questions before you choose
When someone brings me a research proposal, I usually ask the following five questions:
- Is your main goal to test theory or to predict? Testing theory leans toward CB-SEM; prediction leans toward PLS-SEM.
- What is the nature of your constructs: reflective or formative? Formative leans toward PLS-SEM.
- How large is your sample, and how is the data distributed? A very small sample and very non-normal data give a stronger reason for PLS-SEM, but with care.
- How mature is your theory? A theory that is still taking shape suits an exploratory approach better.
- What do your examiner or target journal expect? This is not a scientific reason, but it is a realistic one. If the journal you are aiming for usually uses one approach, understand why and prepare your argument.
The mistake I see most often
It is not choosing LISREL over SmartPLS or the reverse. The most common mistake is being unable to explain why you chose. At a thesis defense, the question “why did you use PLS-SEM?” gets answered with “because my sample is small” or “because it is easier”. A good examiner expects an answer grounded in the research objective, the nature of the constructs, and the characteristics of the data.
The second mistake: skipping the evaluation of the measurement model and jumping straight to reading path coefficients. That is like reading a scale’s result without checking that the scale is accurate. Validity and reliability must be settled first; only then do the relationships between variables mean anything.
The third mistake: making significance the only goal. A model where “everything is significant” is not necessarily meaningful. Look at effect sizes, explanatory power, and whether the relationships make sense theoretically.
Conclusion: choose with reasons, not comfort
SEM versus PLS-SEM is not about who wins. The question is: what do you want to learn from your data, and which tool answers it most honestly? Answer with reasons you can defend, and you will be calmer both in front of your examiner and in front of a journal reviewer.
Two realistic scenarios to build intuition
Scenario A. A lecturer wants to test a well-established theory: service quality affects satisfaction, and satisfaction affects loyalty. She has 500 respondents, all constructs are reflective scales from validated instruments, and she wants to know whether the model “fits” the data and whether it is equivalent for men and women. This is territory for CB-SEM.
Scenario B. A manager wants to understand which factors best predict the success of a customer retention program. The model includes several new constructs with no standard instrument, one of them an index formed from several different indicators (formative), and he has only 120 respondents. The goal is prediction and exploration. This is better suited to PLS-SEM, with the caveat that 120 samples need to be checked for adequacy against the model’s complexity.
Notice that the choice is not decided by the software available, but by the objective, the nature of the constructs, and the data. Whatever you choose, write your reasons in the methods chapter in a few clear sentences. That alone separates research that was thought through from research that was merely run.
Unsure which SEM approach to pick, setting up your model, or interpreting the output? The datalytics.indo team can help with everything from research design and instrument testing, through analysis (CB-SEM and PLS-SEM, path analysis, regression, factor analysis, clustering, and more), to interpretation and writing up the results. Private tutoring in SmartPLS, LISREL/AMOS, and SPSS is also available. Let’s discuss your needs first. We’ll provide a tailored quote based on your project.