Missing data
Outcomes go missing in every trial. Whether that biases the answer depends on why they are missing. See how far each analysis drifts from the truth, and how often its confidence interval still holds the true effect.
Why outcomes go missing
More options
Complete-case bias
-0.09
Average estimate 0.41 against a true 0.5.
Baseline-adjusted bias
+0.01
Average estimate 0.51 against a true 0.5.
| Analysis | Bias | CI holds the truth | Power |
|---|---|---|---|
| No data missing (for comparison) | +0.01 | 94% | 95% |
| Complete cases, unadjusted | -0.09 | 92% | 73% |
| Complete cases adjusted for baseline (as multiple imputation would give) | +0.01 | 94% | 97% |
A 95% interval should hold the true effect 95% of the time; well below that means the analysis misleads.
Outcomes observed
75%
Average 95% CI width (adjusted)
0.51
Spread of the estimates
Amber: complete cases. Teal: adjusted for baseline. The dashed line is the true effect.
For your protocol
With 25% of outcomes missing at random given baseline (more drop-out among control participants with worse baseline values), simulation of a trial with 100 participants per arm and a true difference of 0.5 gave a complete-case estimate of 0.41 (bias -0.09, 95% CI coverage 92%) and a baseline-adjusted estimate of 0.51 (bias 0.01, coverage 94%). The primary analysis will adjust for baseline and use multiple imputation under a missing-at-random assumption, with sensitivity analyses (for example delta-adjustment) for data missing not at random.
How this simulation works
Each simulated trial has a baseline measure that predicts the outcome. Outcomes then go missing completely at random (MCAR), at random given the baseline and the arm (MAR: here, control participants with worse baseline values drop out more), or not at random (MNAR: people with poor outcomes drop out more). The share missing is set by solving for the intercept of a logistic model. The complete-case difference in means is compared with a complete-case regression adjusted for baseline, which under MAR (with only the outcome missing) gives the same answer as multiple imputation using the same variables. Under MNAR no standard analysis is unbiased; only sensitivity analyses can show how much it matters. 1,000 trials per setting.