SEM sample size calculator for a priori planning

How do I calculate sample size for SEM?

Begin with the exact inferential target rather than a generic ratio. Specify whether the design must detect a structural path, loading, indirect effect or model-fit departure; choose the estimator; state the expected effect, alpha and power; and represent non-normality, missing data, clustering and convergence risk where relevant. A model-specific analytic or Monte Carlo power analysis is the strongest route. The calculator above is an earlier planning aid: it displays three heterogeneous approximations so their assumptions can be compared rather than hidden.

How the three displayed estimates work

The Fisher-z result concerns detection of a single correlation under the entered effect, power and alpha; it is not a power analysis for the full SEM. The Westland expression uses the entered indicator-to-latent-variable ratio. The N:q result applies a 10:1 ratio to an approximate parameter count. The largest displayed value is shown for comparison only; taking the maximum does not validate any method for a particular model.

Is there a minimum sample size for SEM or CFA?

No single minimum applies to every SEM or confirmatory factor analysis. A smaller, well-behaved model with strong loadings and complete, approximately normal data may require less information than a complex model with weak effects, categorical indicators, missingness or non-independence. CFA and full structural models therefore need design decisions tied to their own target tests and estimation conditions—not a universal cut-off copied from another study.

Why the 10-times rule is not a power analysis

A parameter-ratio rule can be a transparent rough check, but it does not model the probability of detecting the target effect. It also ignores estimator choice, effect heterogeneity, distribution, missingness, clustering and convergence. Report it as a heuristic only, never as proof that the proposed sample is statistically sufficient.

When should I use Monte Carlo power analysis?

Use Monte Carlo simulation when the actual measurement and structural model matters to the decision—for example, when testing indirect effects, small paths, categorical indicators, non-normal data, missingness, multilevel structure or models at risk of non-convergence. The simulation should reproduce the intended estimator and data-generating assumptions, examine power for the target effect, and report bias, coverage, convergence and admissible-solution rates rather than power alone.

Method version 1.1 · reviewed 31 July 2026. Not represented in the displayed bounds: estimator, degrees of freedom, target parameter or fit statistic, distribution, missingness, non-independence, convergence or model misspecification. Use a model-specific analytic or Monte Carlo design where these features matter; see the semPower package ↗.