Structural equation modeling usually means lavaan syntax, an SEM diagram drawn somewhere else, and a morning spent matching the two. Magic Stat lets you draw the model on a canvas, choose the estimator and run it — the coefficients, the fit indices and the report come out of the same dialog.
SEM estimates a system of relationships at once — measurement (which observed items belong to which latent construct) and structure (which constructs predict which) — instead of fitting one equation at a time. That is what lets it separate measurement error from the relationship you care about, and what makes it appropriate when a construct is measured by several items and you want the construct, not the sum score, in the model.
It is not automatically better than regression. If you have one outcome, observed variables and no latent constructs, an ordinary regression is simpler, more transparent and easier for a reviewer to check. SEM earns its place when there are latent variables, several dependent relationships, mediators or moderators, or when you want a global fit statistic for the whole theory.
The two estimators answer different questions. PLS-SEM is variance-based and oriented to prediction with composite constructs; CB-SEM is covariance-based and gives fit indices you can compare against a hypothesised model. Magic Stat runs both, and states which one produced each number.
Estimator: PLS-SEM · CB-SEM.
Inference: Analytic · Bootstrap (iterations and confidence level set in the dialog).
Group variable: a categorical column for multi-group analysis.
Canvas: save the model as .sem.json, save the diagram as an image, reopen and re-run.
PLS-SEM suits prediction-oriented, exploratory models with composite constructs and modest samples; CB-SEM suits confirmatory testing of a hypothesised model, because it reports fit indices you can judge. If the reviewer expects fit statistics, use CB-SEM; if they expect composite reliability and HTMT, use PLS-SEM.
No. lavaan in R remains the reference for complex CB-SEM — equality constraints, non-standard models, multilevel structures. Magic Stat covers the common reflective model you can draw and gives you the report; if your model needs lavaan's full syntax, use lavaan.
No. The estimators run inside the app. If you already use R comfortably and it serves you, keep it — the point here is the drawing-to-report workflow, not replacing your scripts.
A small number of groups, as in SmartPLS — each group is a full model estimation, so grouping by a near-continuous column is not allowed, and the dialog only lists columns with few levels.
No. The model is estimated on your computer. The only signal the app sends automatically is an anonymous installation counter — no data, no names, no e-mail.