Use FDA Suggestions on Missing Data and Sensitivity Analyses
The U.S. FDA provides guidance and expectations on how to handle missing data in clinical trials, especially in the context of estimands and sensitivity analyses. Below is a detailed summary based on key regulatory documents and practices.
📘 1. Key FDA Guidance Documents
A. FDA (2019): "Estimands and Sensitivity Analyses in Clinical Trials"
-
Focus: Aligning analysis with the trial objective and handling intercurrent events, like treatment discontinuation or dropout.
-
Introduces the estimand framework from ICH E9(R1).
-
Emphasizes sensitivity analysis under plausible MNAR assumptions.
🔗 Link: FDA Estimand Guidance (PDF)
📌 2. FDA Expectations at a Glance
| Topic | FDA View |
|---|---|
| Primary analysis | Can assume MAR (e.g., MMRM) if justified |
| Sensitivity analysis | Must explore MNAR scenarios (not just MAR) |
| Multiple imputation | Acceptable if properly implemented |
| Pattern-mixture models | Encouraged as part of sensitivity analyses |
| Reference-based imputation (CIR, J2R) | Increasingly accepted for conservative MNAR assumptions |
| Pre-specification | All methods (tipping point, MI model, etc.) should be pre-specified in the SAP |
| Use of tipping point | Common and acceptable for visualizing robustness |
| ICHE9(R1) estimand framework | Strongly encouraged; match analysis with estimand strategy |
🧪 3. FDA on CIR and Reference-Based Imputation
The FDA has not officially endorsed one method (e.g., CIR or J2R) over others, but in practice:
-
Reference-based methods like CIR, J2R, and CR are considered valid for sensitivity analyses under MNAR.
-
They are increasingly used in late-phase confirmatory trials and regulatory submissions.
-
FDA statisticians have accepted CIR/J2R approaches in NDAs/BLAs, particularly in placebo-controlled trials when dropout is informative.
They are particularly useful under the "treatment policy" or "hypothetical" estimand strategies.
✅ CIR is not expected to be the primary analysis, but should be included as part of a robust sensitivity analysis suite.
✅ 4. What FDA Recommends You Do
Here’s what you should include in your Statistical Analysis Plan (SAP) or trial design documents:
A. Clearly Define the Estimand
-
What treatment effect are you estimating?
-
How are intercurrent events (like dropout or rescue therapy) handled?
B. Pre-specify Primary and Sensitivity Analyses
-
Use MMRM under MAR as the primary if appropriate
-
Add at least one or two MNAR sensitivity analyses:
-
Pattern-mixture model
-
CIR or J2R
-
Tipping point
-
C. Justify Imputation Assumptions
-
Why is MAR plausible (or not)?
-
What is the clinical reasoning behind CIR/J2R shift values or patterns?
D. Use Rubin's Rules for MI and match analysis model to primary
-
Ensure consistency of model form (e.g., MMRM in all imputations)
-
Document the imputation model and convergence diagnostics
🧾 5. Example Language from FDA-Reviewed SAP
"The primary analysis will be performed using a mixed model for repeated measures (MMRM) under the assumption that missing data are missing at random (MAR). Sensitivity analyses will include pattern-mixture models with delta-adjusted multiple imputation and reference-based multiple imputation using the copy increments in reference (CIR) and jump-to-reference (J2R) approaches, to explore departures from the MAR assumption."
📎 Summary
| Method | FDA Status | Use Case |
|---|---|---|
| MMRM | ✅ Accepted | Primary analysis under MAR |
| Tipping point | ✅ Encouraged | Sensitivity to departure from MAR |
| Pattern-mixture models | ✅ Recommended | Sensitivity analyses (MNAR) |
| Reference-based MI (CIR, J2R) | ✅ Increasingly used | Conservative sensitivity scenarios |
| Complete-case / LOCF | ❌ Not recommended | Biased, outdated methods |
The copy increments
in reference (CIR) method is a reference-based sensitivity analysis for assessing
the robustness of the primary analysis to deviation from the MAR assumption. This method assumes that any benefit
gained from previous treatment will be retained, but patients will progress as
if they were in the reference group (placebo group) after withdrawal from the
study.
With this method,
the missing data after a patient’s withdrawal from the study for patients who
receive the active treatment will be imputed based on data from the placebo
group. Specifically, for a patient on active treatment who withdraws early, his
mean trajectory after early withdrawal is assumed to be parallel to the mean
trajectory of the placebo group. For a patient on placebo who withdrew early,
his post-withdrawal profile will be imputed following the MAR principle.
After all missing
data has been imputed, the MMRM analysis will be conducted as described in the
primary analysis. The analyzed results from the m (=100) imputed datasets will
be combined based on Rubin’s rules assuming the
statistics estimated from each imputed dataset are normally distributed.
1️⃣ Monotone Missingness
Once a subject has a missing value, all subsequent values are also missing.
🔁 Example:
| Subject | Week 0 | Week 12 | Week 24 | Week 48 |
|---|---|---|---|---|
| A | 10 | 12 | 13 | NA |
| B | 8 | NA | NA | NA |
✅ Subject B has monotone missingness starting at Week 12
✅ Subject A has monotone missingness starting at Week 48
2️⃣ Non-Monotone Missingness
A subject has missing data at intermediate time points, but has data again at later visits.
🔁 Example:
| Subject | Week 0 | Week 12 | Week 24 | Week 48 |
|---|---|---|---|---|
| C | 9 | NA | 11 | 12 |
❌ Subject C has non-monotone missingness — the Week 12 data is missing, but Week 24 and 48 are observed.
📊 Why It Matters
| Feature | Monotone | Non-Monotone |
|---|---|---|
| Easier to handle? | ✅ Yes | ❌ More complex |
| Common in trials? | ✅ Yes (due to dropout) | ✅ Yes (intermittent missing) |
| Imputation methods | Simple regression-based MI | Requires iterative methods (FCS / MCMC) |
| Tipping point models | Usually assume monotone | Harder to implement |
| Sensitivity analyses | Often easier to design for monotone | Need special care |
🧠 In Practice
Monotone:
-
Usually caused by dropout, discontinuation, or death
-
Makes modeling assumptions easier
-
Works well with delta-adjusted imputation, reference-based methods, etc.
Non-Monotone:
-
Can be due to missed visits, data entry errors, intermittent noncompliance
-
Requires more flexible models, like Fully Conditional Specification (FCS) in
miceorPROC MI
🧪 Multiple Imputation Examples
| Missing Pattern | SAS Method | R Method |
|---|---|---|
| Monotone | PROC MI with MONOTONE method (e.g., REG, LOGISTIC) | mice() with method = "norm" and monotone predictor matrix |
| Non-Monotone | PROC MI with FCS or MCMC | mice() with FCS (method = c("norm", "logreg", ...)) |
📘 Summary Table
| Aspect | Monotone Missing | Non-Monotone Missing |
|---|---|---|
| Definition | Missingness persists once it starts | Missingness appears in a non-sequential pattern |
| Example | Dropouts | Missed visits, intermittent |
| Imputation strategy | Easier, sequential regression | Requires FCS or joint modeling |
| Software in SAS | MONOTONE REG, LOGISTIC | FCS, MCMC |
| Software in R | mice() with method="norm" | mice() with full FCS setup |
| Regulatory sensitivity use | Very common (CIR, tipping) | Less common, more complex |
🧭 Practical Tip
If you have a mostly monotone pattern (e.g., 80% of missing due to dropout), it’s reasonable to treat it as monotone for sensitivity analysis (e.g., tipping point or CIR). But for imputation under MAR, you should respect the actual non-monotone structure and use FCS.
Comments
Post a Comment