From: A tutorial on sensitivity analyses in clinical trials: the what, why, when and how
Scenario | Sensitivity analysis options |
---|---|
Outliers | - Assess outlier by z-score or boxplot |
- Perform analyses with and without outliers | |
Non-compliance or protocol violation in RCTs | Perform |
- Intention-to-treat analysis (as primary analysis) | |
- As-treated analysis | |
- Per-protocol analysis | |
Missing data | - Analyze only complete cases |
- Impute the missing data using single or multiple imputation methods and redo the analysis | |
Definitions of outcomes | - Perform analyses on outcomes of different cut-offs or definitions |
Clustering or correlation | - Compare the analysis that ignores clustering with one primary method chosen to account for clustering |
and multi-center trials | |
- Compare the analysis that ignores clustering with several methods of accounting for clustering [10, 11] | |
- Perform analysis with and without adjusting for center | |
- Use different methods of adjusting for center [12] | |
Competing risks in RCTs | - Perform a survival analysis for each event separately |
- Use a proportional sub-distribution hazard model (Fine & Grey approach) | |
- Fit one model by taking into account all the competing risks together [13] | |
Baseline imbalance | Perform: |
- Analysis with and without adjustment for baseline characteristics | |
- Analysis with different methods of adjusting for baseline imbalance. e.g. Multivariable regression vs. propensity score method | |
Distributional assumptions | Perform analyses under different distributional assumptions |
- Different distributions (e.g. Poisson vs. Negative binomial) | |
- Parametric vs. non-parametric methods | |
- Classical vs. Bayesian methods | |
 | - Different prior distributions |