The difference between MMRM with vs without random effects

 

🧠 Short Answer

Model Type	Random Effect Included?	Type of Model	Assumes Subject-specific Trajectories?	Typical Use
MMRM	❌ No	Marginal model	❌ No	Regulatory trials (e.g. FDA)
Linear Mixed Model (LME)	✅ Yes	Conditional model	✅ Yes	Academic studies, hierarchical modeling

---

🔍 MMRM = No Random Effects

✅ Standard MMRM (FDA-favored)

• Marginal model: estimates population-average effects.

• Uses a repeated statement in SAS (or nlme::gls() in R).

• Models the within-subject covariance directly (e.g., UN, AR(1), CS).

• Assumes no subject-specific random intercepts/slopes.

➕ Pros:

• Very flexible in modeling within-subject correlations.

• Doesn't assume normally distributed random effects.

• Performs well under MAR with dropout.

• Standard in late-phase clinical trials and regulatory submissions.

➖ Cons:

• Can become unstable with too many timepoints (e.g., unstructured covariance).

• Cannot model subject-level variation (e.g., random slopes).

📘 Example in SAS:

proc mixed data=adsl;
  class id visit trt;
  model change = trt visit trt*visit / solution;
  repeated visit / subject=id type=un;  /* ← within-subject covariance only */
run;

---

🔍 LME = Random Effects Included

✅ Linear Mixed-Effects Model (aka Hierarchical Model)

• Conditional model: estimates effects within individuals.

• Uses a random statement for subject-level intercepts/slopes.

• Residuals assumed i.i.d. after accounting for random effects.

• Models the intra-subject variability via random intercepts/slopes.

➕ Pros:

• Models subject-specific deviations (e.g., individual growth curves).

• Better for small sample sizes, nested/hierarchical structures.

• Often used in academic and psychometrics research.

➖ Cons:

• Assumes random effects are normally distributed.

• Not ideal when missing data are not MCAR, especially with dropout.

• Not preferred in regulatory settings (can bias treatment effect estimates under MAR with dropout).

📘 Example in SAS:

proc mixed data=adsl;
  class id visit trt;
  model change = trt visit trt*visit / solution;
  random intercept / subject=id;  /* ← random effect: subject-specific intercept */
run;

---

📈 Key Differences

Feature	MMRM (No Random Effect)	Mixed Model (With Random Effect)
Type of model	Marginal (population-level)	Conditional (subject-level)
Target estimand	Mean difference across population	Individual trajectories
Handles dropout?	✅ Yes, under MAR	❌ Biased if dropout is informative
Within-subject correlation	Modeled via repeated covariance	Modeled via random effects
Between-subject variability	Not explicitly modeled	✅ Modeled via random intercepts/slopes
FDA/EMA preference	✅ Yes	❌ Generally not preferred for primary analysis
Example use	Phase III trials	Longitudinal growth, academic studies

---

🧪 Which Should You Use?

If You...	Use
Need regulatory-compliant population-level estimates	✅ MMRM
Want to model individual-level change over time	✅ Random effects (LME)
Have monotone missing data and want unbiased inference under MAR	✅ MMRM
Want to model hierarchical structure (e.g., patients nested in sites)	✅ Mixed model with random intercept for site

---

🧭 Bottom Line

• MMRM is the gold standard for continuous longitudinal outcomes in clinical trials.

• Random effects models are powerful but not robust to informative dropout—use with caution for regulatory work.