How to Choose the Right Correlation-Corrected Statistic: Model Hierarchy for Repeated Measures (Best → Fallback)
Clinical Epidemiology ResearchUniqcret doctor knowledgesData Analytics or StatisticsMethodology and Research Design
On this page
📊 Model Hierarchy for Repeated Measures (Best → Fallback)
| Rank | Model | Description |
| 🥇 1 | Conditional Model — Multi-level — Random Effect | Subject-specific; captures both random intercepts and slopes |
| 🥈 2 | Conditional Model — Multi-level — Fixed Effect | Subject/group-specific using dummy variables; no generalization |
| 🥉 3 | Marginal Model — Single-level — Model-Based Variance | Population-averaged; requires correlation structure assumption |
| 🪙 4 | Marginal Model — Single-level — Empirical (Robust SE) | Uses sandwich SE; structure-free fallback |
🔁 Downgrade Pathway (When Data is Limited)
| Limitation | Downgrade to... | Reason |
| No group ID (e.g. id, cluster) | ❌ Can't use Conditional → Use Marginal | Can't define within-subject correlation |
| Few groups (<5) | ❌ Random → ✅ Fixed Effect | Random effects may overfit |
| Insufficient data for random slopes | ❌ Random slopes → ✅ Random intercept or Fixed | Model may not converge |
| Unknown correlation structure | ❌ Model-based → ✅ Empirical (robust SE) | Robust to mis-specification |
⚔️ 1. Random Effect vs Fixed Effect (Conditional Multilevel)
| Criterion | Fixed Effect | Random Effect |
| Uses dummy variables? | ✅ Yes | ❌ No (learns from distribution) |
| Can predict new/unseen groups? | ❌ No | ✅ Yes |
| Requires all groups in model? | ✅ Yes | ❌ No |
| Learns group-level variance? | ❌ No | ✅ Yes (estimates variance components) |
| Allows random slopes? | ❌ No | ✅ Yes (e.g., ` |
| Suitable for few groups? | ✅ Yes | ❌ Not stable with few clusters |
| Stata commands | xtreg, fe, reg outcome i.group | `mixed outcome ... |
⚔️ 2. Fixed Effect vs Marginal Model-Based
| Criterion | Fixed Effect | Marginal (Model-Based) |
| Focus on individual/group effect? | ✅ Yes | ❌ No (population-averaged) |
| Can predict individual response? | ✅ Limited to included groups | ❌ Not designed for this |
| Command in Stata | xtreg, fe, areg, reg i.group | xtgee ... corr(exchangeable) |
| Population-level interpretation? | ❌ Not primary aim | ✅ Yes |
| Needs large sample for stability? | ❌ Less dependent | ✅ More sensitive to cluster size |
| Suitable for few clusters? | ✅ Yes | ❌ Risky, may produce unstable estimates |
⚔️ 3. Marginal Model-Based vs Empirical (Robust SE)
| Criterion | Model-Based Variance | Empirical (Robust SE) |
| Requires correlation structure? | ✅ Yes (e.g., AR1, exchangeable) | ❌ No |
| Sensitive to mis-specification? | ❌ Yes | ✅ No (SE remains valid) |
| Uses sandwich estimator? | ❌ No | ✅ Yes |
| Higher power when correct? | ✅ Yes | ❌ Lower power if model structure unknown |
| Best for... | Known structure, more clusters | Small samples or uncertain correlation |
| Stata commands | xtgee ... corr(ar1) | xtgee ... vce(robust) |
🔚 Final Decision Strategy
Start with:
- Mixed Model (Random Intercept/Slope)→ if not feasible due to convergence/small N
- Fixed Effects Model→ if still not feasible
- GEE with correlation structure→ if structure is unclear or unreliable
- GEE with robust (Empirical) SE ← safest fallback
Let me know if you want side-by-side Stata syntax, visual simulations, or a clinical use case example.
0