The fastsem engine was built to handle workloads where existing SEM tools either cannot run the model at all or are impractically slow. The matrix below summarises the differences across the features that motivated its development.
Feature matrix
| Feature | fastsem | Mplus | OpenMx | lavaan | Stan |
|---|---|---|---|---|---|
| Native IV / MR (Gen-SEM) | ✓ | ~ | ✓ | ✗ | ~ |
| State-space / Kalman (SSM) | ✓ | ✗ | ✓ | ✗ | ~ |
| Ordinal thresholds + FIML | ✓ | ✓ | ✓ | ~ | ~ |
| GPU acceleration (OpenCL) | ✓ | ✗ | ✗ | ✗ | ~ |
| R interface | ✓ | ✗ | ✓ | ✓ | ✓ |
| Multi-group + cluster-robust SEs | ✓ | ✓ | ✓ | ✓ | ✗ |
| PLINK binary + VCF streaming | ✓ | ✗ | ✗ | ✗ | ✗ |
✓ = supported natively, ~ = partial /
requires manual work, ✗ = not available.
When fastsem is the right tool
- Genomic SEM at scale. Streaming PLINK/VCF input plus warm-started per-SNP fits make genome-wide SEM scans tractable on a single workstation.
- Mendelian randomisation / IV inside an SEM. Native IV support via lavaan-style regression syntax, with no separate two-stage code.
- Ordinal indicators with missing data. Polychoric correlations + DWLS, combined with FIML for incomplete patterns.
- State-space / Kalman models. GPU-accelerated Kalman filter for latent-trajectory and continuous-time models.
- Definition variables. Per-row parameters fit via FIML, useful for twin/family designs and individually varying loadings.
When another tool may be a better fit
- Bayesian SEM with custom priors: use Stan or OpenMx’s MCMC backend until fastsem’s NUTS sampler lands (see the roadmap).
-
Mature ecosystem of plotting /
semTools-style diagnostics: lavaan has the deepest set of post-processing utilities.
Honest caveats
The comparison above reflects each tool’s out-of-the-box capability. With sufficient custom code most tasks can be expressed in any of these systems; the matrix is about what comes for free.
See the benchmarks article for wall-clock numbers on representative workloads.